Talk by Remy Guercio (Tailscale) examining a year of AI tool lock-in and how teams can escape it while demonstrating ROI as unlimited AI budgets end. He breaks an internal AI deployment into four layers—models, data connectors (MCP, CLIs, APIs), interface, and sandbox—and argues an AI gateway keeps them swappable. He presents Tailscale's Aperture gateway: identity derived from the Tailscale network instead of shared API keys, per-model budgets, dual LLM and MCP gateway roles, centralized authentication for tools like Cursor, Claude Code, and Codex, and logs for diagnosing agent behavior. # Channel Presentation Observations: Remy Guercio, Tailscale ## Source Scope - **Evidence base:** One video transcript and title. No curated channel metadata was supplied. - **Format:** Conference talk, delivered as the closing session of an event described as the "AI engineer world's fair." - **Limitation:** These observations describe this video only. They do not establish channel-wide patterns. --- ## 1. Explicit Self-Description (Creator Claims) The speaker describes his own role and employer directly: > "My name is Remy. I am on the strategic projects team at Tailscale. Uh and more specifically I actually work on our AI gateway aperture" - **Role:** Strategic projects team member at Tailscale. - **Product area:** Works on the AI gateway product called **Aperture**. - **Event context:** Describes the session as "the last talk on this stage" at the event. The speaker also describes his vantage point as a basis for his perspective: > "I work on an AI gateway. We use our AI gateway. I get to see all of the usage from the AI gateway." **Note:** The speaker presents this vantage point as a source of insight, but the usage data itself is not shown in the transcript. --- ## 2. Intended Audience (Observed from Content) The transcript addresses an audience of **organizational AI deployers**, not general viewers. Evidence includes: - **Internal deployment framing:** The speaker says he is not talking about "a hey, we've built this new AI feature in our app. I mean... internal to your organization." - **Organizational roles referenced:** "finance is talking to you," "the marketing team," "the engineering team." - **Technical audience:** Discussion of MCP, CLI, APIs, sandboxes, and token context windows assumes technical familiarity. > "We're just going to consolidate on a single vendor maybe two vendors" This is a description of a common scenario the speaker reports, framed as advice to avoid. --- ## 3. Claimed Scope and Thematic Focus The talk claims to address how to avoid **AI vendor lock-in** and improve return on AI investment. ### Core Concepts Introduced - **Token maxing:** Describes the last 12 months of unconstrained token use. The speaker explicitly separates this from its negative connotation. - **ROI maxing:** Proposed successor framing. The speaker states it means "not just cutting costs... actually proving some sort of value." - **Four components of internal AI deployment:** 1. LLM(s) 2. Data connector layer (MCP, CLI, APIs, or file reads) 3. Interface (chat, mobile, Slack, and similar) 4. Sandboxes (execution environments) ### Thesis Statement > "ROI maxing really is all about maximizing choice to prevent lock in" --- ## 4. Product Promotion and Balance The talk features a specific product (Aperture) with demonstrated feature claims: - Centralized LLM provider configuration and budget control - MCP server configuration with role-based access (for example, marketing versus engineering) - Identity propagation via Tailscale's identity-based mesh network - A planned orchestration layer for sandboxes - A planned pass-through OAuth feature for interfaces such as Open WebUI The speaker also includes a hedge that frames the recommendation more broadly: > "even if you don't use Aperture I do highly recommend you start considering an AI gateway" **Observation:** The talk combines product demonstration with general industry guidance. The speaker explicitly states that the recommendation applies beyond his company's product. --- ## 5. Experiential and Anecdotal Claims The speaker makes several claims based on his professional interactions. These are presented as first-hand observations and are not independently verified in the transcript: - **Scale of conversations:** "over the last 6 months, we've been talking to... hundreds and hundreds of folks about their AI spend" - **Usage patterns:** "sessions that are hundreds or thousands of messages long where nobody clears the context" - **Cost magnitude:** "tens of millions of tokens of a session" - **Agent behavior:** Agents "spiraled out of control," and cheaper models are "usually not great at stopping themselves" > "they're usually not great at stopping themselves. Uh but they are great at analyzing where they went wrong after the fact." --- ## 6. Forward-Looking Claims The speaker makes predictions about near-term industry direction: > "I think this is kind of peeking at what we're going to see the next 12 months" These are speculative and framed as the speaker's own view. --- ## 7. Delivery and Stylistic Patterns Observable from the transcript (text-based, so visual and tonal evidence is unavailable): - **Conversational register:** Frequent filler words, self-corrections, and informal phrasing ("you know," "uh," "right?"). - **Rhetorical questions and direct address:** "how many people like have an AI gateway... Yeah." - **Mixed structure:** Problem framing, then named solution categories, then product walkthrough, then closing summary. - **Self-referential humor and light framing:** Calls the four-component breakdown "a reminder here." --- ## 8. Competitor and Ecosystem References The speaker names several third-party tools and vendors: - Claude Code, Codex, Cursor, Open WebUI, OpenClaw - Anthropic, OpenAI, Bedrock, Vercel, Entra, Okta - Named models in passing (for example, "GLM 5.2") These references are used as examples of the lock-in landscape and of integration targets. --- ## Summary of Verified Versus Unverified Content | Category | Status in Transcript | |---|---| | Speaker name, employer, and role | Explicitly self-stated | | Product being promoted | Explicitly stated and demonstrated through description | | Event context | Explicitly stated | | Audience (organizational AI deployers) | Inferred from content, not stated directly | | Scale of customer conversations | Creator claim, unverified | | Usage statistics (token counts, session lengths) | Creator claim, unverified | | Forward predictions | Speculative creator claim | | Channel-wide patterns | Not assessable from this single video | # Delivery Observation — Remy Guercio, "An AI Future Without the Lock-In" ## Overall Register The delivery is **conversational and informal**, closer to a hallway conversation than a formal keynote. This is visible in the opener: > "Well, hi folks. Uh, thanks for joining me here for the last talk on this stage uh of the uh AI engineer world's fair" High filler-word density ("uh," "um," "you know," "right") appears throughout, suggesting an unscripted, spoken-style delivery rather than a rehearsed read. ## Recurring Vocabulary & Phrases Several terms repeat at a frequency high enough to count as stable verbal habits: - **"Right?"** as a tag question / rhetorical confirmation device — used constantly to pull the audience along: > "token maxing is cool but right it has it has problems right there's obvious problems" - **"You know"** as a connective filler, often stacked: > "you know what it means or you know what it's going to take uh to have an AI future without the lockin" - **"-maxing" coinages** as a framing device — the talk's central rhetorical pattern: > "we're not going to be token maxing we're going to be ROI maxing" This is introduced, defined ("I don't just mean cutting costs right I mean actually proving some sort of value"), and revisited at the close ("ROI maxing really is all about maximizing choice"). - **Self-aware hedging on vendor interest**: repeatedly acknowledges the sales context before making the pitch: > "maybe no surprise that I'm here to talk a little bit about why an AI gateway uh will help you" > "I don't just mean necessarily using our AI gateway I think it's nice but you know just any AI gateway" ## Teaching Habits - **Enumerated frameworks**: introduces a four-part structure and explicitly numbers it: > "there's four things, right? There's the LLM or LLMs... there's sort of the data connector layer... there's the interface... and then the last one is sandboxes" He then walks each layer one at a time against the product, and recaps it: "just reminder components of a you know an internal you know AI deployment." - **"How not to" sequencing**: teaches by negative example before the positive one: > "but first I want to talk about how not to address the problem." - **Concrete anecdotes as evidence**: illustrative workplace scenarios rather than abstract claims: > "I get, you know, uh, I get an anthropic bill every day." > "the marketing team, you know, they get access to Google Calendar and Gmail, but the engineering team, they'll also get access to GitHub" - **Audience-directed asides and show-of-hands moments**: > "how many people like have an AI gateway, have deployed an AI gateway" > "I'm sure I could poll everybody here and would have a different everybody would have a different definition" ## Communication Patterns - **Contrast / negation-then-correction** is a frequent explanatory move ("I don't mean X, I mean Y"): > "I don't mean that in the kind of you know negative sense... I mean that in the sort of you know it's been a year of trying to figure out..." - **Hedged generalizations** soften strong claims: > "that is not not a great idea. uh you know I might seem obvious sitting here it's maybe not a great idea" - **Playful, casual humor** aimed at shared practitioner pain: > "you can tell exactly when an MCP server was built by exactly how they decided to do off. uh yeah, it's it's a great uh you know kind of archaeology lesson" - **Humor at model behavior's expense**: > "It just kind of spiraled out of control. Uh maybe happens less now. Definitely used to happen quite a bit." - **Signposted structure** with casual transitions ("So um all right," "and then the last thing," "So um that's it uh for my talk just short and sweet"). ## Isolated vs. Stable Observations - **Stable (high repetition)**: "right?" tag questions, "you know" fillers, the "-maxing" framing, enumerated component walkthroughs, and self-aware vendor hedging all recur throughout. - **Thin evidence (single instances)**: the humor about MCP archaeology and the "spiraled out of control" joke appear once each; they suggest a joking register but cannot be confirmed as recurring habits from this transcript alone. The "let a thousand flowers bloom" metaphor is likewise a one-off flourish. ## Uncertainty Notes - The transcript's filler-heavy style may partly reflect transcription artifacts; only patterns visible in the text itself are reported here. - Slide/visual references (e.g., "as you can see here") are audible but the visuals are not available, so claims about demo structure are limited to what is spoken. # Video Position Record: "An AI Future Without the Lock-In" — Remy Guercio, Tailscale ## Speaker Identification - **Remy Guercio** — Strategic Projects team at Tailscale; works on Tailscale's AI gateway, **Aperture**. (No curated speaker metadata beyond transcript self-identification was supplied.) --- ## Contextual Framing: The Last 12 Months as "Token Maxing" > "It can really be described quite a bit by the phrase token maxing." Guercio characterizes the prior year positively: million-token context windows, context from anywhere (MCP, CLI-based coding tools), enabling **truly agentic coding** and the emergence of other agentic use cases. He explicitly notes he doesn't mean the phrase in the negative sense it may carry outside AI-enthusiast circles. ### Problems Identified with Token Maxing - **Cost**: "It's an extremely expensive endeavor"; the era of "unlimited AI budget" has ended "here in the last few months." - **Wasted tokens by default**: From his gateway visibility, he reports sessions of "hundreds or thousands of messages long where nobody clears the context," with "tens of millions of tokens" per session and cache maxed the entire time. - **Vertical integration / lock-in**: Toolchains built "all the way from the model up to the top" (e.g., Claude Code, Codex). --- ## Claim: Vendor Consolidation Is the Wrong Response - Over the last six months, Guercio says he has spoken to "hundreds and hundreds of folks" about AI spend; common pattern: daily Anthropic and Codex bills, uncertainty about usage. - The instinctive fix — **consolidating on one or two vendors** — he calls "not a great idea," while acknowledging it's a "pretty natural response" when finance is pressuring teams and panic sets in. - Qualification preserved: he concedes this may seem obvious in hindsight but not "when you're in that panic zone." ## Claim: The Next 12 Months Will Be "ROI Maxing" > "We're not going to be token maxing, we're going to be ROI maxing." - He defines ROI maxing as **not just cutting costs**, but "actually proving some sort of value" and "opening things up." - Reasoning: if the only option is "two frontier models and their entire stacked ecosystem," that fixes a cost floor that "makes certain things untenable" — you cannot expand AI usage across the company. --- ## Framework: Four Components of an Internal AI Deployment Guercio explicitly excludes customer-facing AI features; this is about **internal** deployment: 1. **LLM(s)** — the model layer 2. **Data connector layer** — "MCP for a lot of people, but in reality... CLI or APIs or really just any way that you get data, could be reading a file" 3. **Interface** — traces the evolution: chat → IDEs with chat → mobile → "Claude tag" / agents via Slack; notes it's a breakable-out component 4. **Sandboxes** — "just the environment in which the agent runs" **Core thesis**: ROI maxing means "maxing every opportunity... to make a choice at one of those layers." He qualifies this — you don't need a bespoke choice at every layer, but you must be *able* to choose at any layer, at any time. Choices may differ by team: he says he doesn't "totally imagine" an engineering team coding via Slack, but a marketing team might work entirely in Slack threads. --- ## Claim: An AI Gateway Enables Choice at Every Layer - Position: "Maybe no surprise... I'm here to talk a little bit about why an AI gateway will help you do that," with the explicit qualification: "I don't just mean necessarily using our AI gateway... just any AI gateway is important for this." - A gateway sits between agents/interfaces and all endpoints (LLM and data), enabling **control and monitoring** — security, cost, and behavioral analysis. - On debugging: agents "are usually not great at stopping themselves," but "are great at analyzing where they went wrong after the fact" — logs let you point the agent back at what it did and ask what happened. He notes spiraling happened less now than previously, and more with cheaper models. ### Aperture-Specific Design Claims - **Identity via Tailscale**: Built on Tailscale, an identity-based mesh network on WireGuard. Every connection carries the identity of the machine or user (tags like a PR-review bot with a node ID, users, or synced groups from Entra/Okta). Built with the open-source **TSNet** library. - **No synthetic-key problem**: Typical gateways require "trading some keys for a bunch of other keys"; with Aperture, "just add your API key and go" — the gateway knows who/what you are and grants per-identity access and budgets (e.g., "unlimited budget with a cheap model and an extremely narrow budget with a frontier model"). - **"AI gateway" vs "LLM gateway"**: He says he "very specifically chose" the broader term because Aperture is both an LLM gateway *and* an MCP gateway, and can do OAuth injection for standard APIs, re-exposing them as MCPs. - **MCP fragmentation**: "There's a standard, but everybody follows a different version of that standard... you can tell exactly when an MCP server was built by exactly how they decided to do auth" — which he calls an "archaeology lesson." - **Lower switching friction**: Authentication happens at the gateway, so users of Cursor, Claude Code, Codex, or new agent frameworks "just hit that endpoint" — no re-authing per tool. He argues "all those little points of friction" are what stop people from experimenting with cheaper or faster alternatives. - **Interface layer**: Aperture ships a built-in chat UI but exposes underlying APIs so users can build Slackbots, mobile/voice interfaces, etc. Upcoming: **pass-through OAuth** so third-party UIs (Open WebUI, LibreChat-style) don't lose user/machine identity — he notes such setups currently make "everything... just coming from the system." - **Sandbox layer**: He judges this the layer where "folks are doing pretty good," in a "let a thousand flowers bloom" phase, precisely because "everybody would have a different definition of exactly what a sandbox is." Aperture is adding a **sandbox orchestration layer**, using the term "AI gateway" "intentionally very loosely." --- ## Closing Recommendation - "ROI maxing really is all about maximizing choice to prevent lock-in." - Even without Aperture: "I do highly recommend you start considering an AI gateway" — valued not only for provider switching but for **understanding what agents are doing**, citing individuals running agents like OpenClaw on their tailnet in the background who discover it "did something it was not supposed to do." - Final reframing: the value of a gateway is "not just lock-in right, it's actually just experimentation and understanding as a whole." --- ## Noted Tensions (Preserved, Not Smoothed) - Guercio acknowledges the consolidation instinct he criticizes is a **natural panic response**, not mere error. - He repeatedly discloses his vendor interest ("maybe no surprise," "I work at Tailscale, I work on an AI gateway") while explicitly extending the recommendation to *any* gateway. - He calls the loose use of "AI gateway" "kind of nebulous" while defending the deliberate broadening of the term. - His characterization of the token-maxing era is explicitly **two-sided**: "quite cool" and enabling, yet expensive and wasteful by default. # Attribution Observation: "An AI Future Without the Lock-In" **Video ID:** `01a15ca6-ad7c-71b3-8e68-562b69479ad3` **Title (as supplied):** An AI Future Without the Lock-In, Remy Guercio, Tailscale --- ## 1. Speakers and Roles ### 1.1 Primary Speaker - **Name:** Remy Guercio (self-identified as "Remy" in the transcript) - **Stated affiliation:** Tailscale, "strategic projects team" - **Stated responsibility:** Works on Tailscale's **AI gateway, Aperture** - **Role in this video:** Sole speaker. Presents as the conference talk's speaker and provides first-person accounts of work at Tailscale. - **Setting:** Introduces the talk as the "last talk on this stage" at the "AI engineer world's fair this year." ### 1.2 Other Voices - **None identified.** The transcript contains no questions, audience responses, moderators, or other speakers. - Audience members are addressed directly ("folks," "everybody here") and asked for a show of hands, but no audience reply is transcribed. --- ## 2. Speaker-Reported Claims About Their Own Experience These are first-person claims. They are recorded as the speaker's statements and are not verified against any other source. - **Customer conversations:** Says they have talked with "hundreds and hundreds of folks" about AI spend and usage over the last 6 months. - **Gateway telemetry:** Says they see usage from Tailscale's own AI gateway, including sessions of "hundreds or thousands of messages" and "tens of millions of tokens." - **Personal anecdote:** Describes receiving "an anthropic bill every day" and "a codeex bill every day." - **Hypothetical or example scenarios:** Describes a team that "consolidate[s] on a single vendor" and a user running openclaw on a tailnet. The transcript does not establish whether these are actual cases or illustrations. --- ## 3. Referenced Sources and Third-Party Entities ### 3.1 Vendors and Products Named | Category | Named Items | |---|---| | **AI model providers** | Anthropic, OpenAI, Vercel, AWS Bedrock (as "bedrock"), Codex ("codeex") | | **Coding and agent tools** | Claude Code ("cloud code"), Codex, Cursor, openclaw | | **Identity providers** | Entra, Okta ("Octa") | | **Productivity apps (as MCP data sources)** | Google Calendar, Gmail, GitHub | | **Chat interfaces (open source / third party)** | Open WebUI, LibreChat (transcribed as "Libra like") | | **Interfaces named as examples** | Slack, mobile apps | | **Models named** | GLM 5.2 (cheap-model example), unnamed "frontier models" | | **Protocols and libraries** | MCP (Model Context Protocol), WireGuard, TSNet (described as open source), CLI, APIs | ### 3.2 Event Referenced - **AI engineer world's fair** (the speaker's "this year" event, where the talk was delivered) ### 3.3 Specific Claims Attributed to Others - **General industry trend:** Describes "a lot of the locking that we've seen… over the last 12 months" and "vertically integrated tool chains" built by Claude Code and Codex. These are the speaker's characterizations. - **Customer behavior:** Says people "come with one kind of solution in mind" (consolidating on a vendor). Attributed to unnamed customers. - **Finance involvement:** Says "finance is talking to you" in describing customer pressure. Attributed to unnamed customers' organizations. - **Model behavior:** States cheaper models are "usually not great at stopping themselves" but "great at analyzing where they went wrong after the fact." Presented as the speaker's observation without a cited source. - **Standards fragmentation:** States "everybody follows a different version of that standard" regarding MCP. Presented without a citation. ### 3.4 Explicit Citations - **None.** No URLs, papers, reports, or dated publications are cited in the transcript. --- ## 4. Promotional Content > **Promotion flag: the talk is substantially promotional for Tailscale and its product Aperture.** ### 4.1 Promoted Products and Features - **Aperture** (Tailscale's AI gateway): Described at length as covering LLM routing, MCP gateway functions, a built-in chat UI, identity pass-through, and a planned sandbox orchestration layer. - **Tailscale** (identity-based mesh network built on WireGuard): Described as the foundation for Aperture's identity handling. - **TSNet** (open source library): Described as the library used to build Aperture. ### 4.2 Forward-Looking Promotional Statements - "We're actually going to be adding" a full pass-through OAuth capability (described as "coming very soon"). - "We're actually going to be adding kind of an orchestration layer as well for sandboxes." - "The whole point is you're going to be able to do that with Aperture." These are roadmap statements by the speaker. They are not confirmed by any other source in the record. ### 4.3 Direct Calls to Action - "I do highly recommend you start considering an AI gateway," with the caveat "even if you don't use Aperture." ### 4.4 Promotional Segments (Approximate by Content, No Timestamps Available) - **Segment A:** Aperture overview and components (LLM, data, interface, sandbox layers). - **Segment B:** Tailscale identity model and TSNet. - **Segment C:** LLM provider configuration in Aperture. - **Segment D:** MCP server and access-control configuration in Aperture. - **Segment E:** Built-in chat UI and planned identity pass-through. - **Segment F:** Sandbox orchestration plans. - **Closing:** Recommendation to adopt an AI gateway, with the Aperture reference. --- ## 5. Provenance and Availability ### 5.1 Available Evidence Channels - **Transcript:** Present and used for this record. - **Title:** Present ("An AI Future Without the Lock-In — Remy Guercio, Tailscale"). ### 5.2 Unavailable or Not Supplied - **Timestamps:** Not supplied. Segment references above are content-based only. - **Curated metadata beyond title:** No description, upload date, channel name, or speaker bio supplied. - **Slides or visual materials:** Not supplied. The transcript references a visual ("if I had this expanded") and audience gestures, but these cannot be verified from text. - **Comments:** Not supplied. Comment availability is unknown. - **External verification of speaker identity or affiliation:** Not performed. Affiliation with Tailscale rests on the speaker's own statements in the transcript. - **Source citations for any factual claim:** None available in the transcript. ### 5.3 Transcript Limitations - The transcript is auto-style and contains apparent mis-transcriptions of product and vendor names (for example, "cloud code," "codeex," "Octa," "Libra like," "openclaw," "versel," "GLM 5.2"). Names are recorded above as they appear, with likely intended forms noted where clear. - Some passages are repetitive or contain filler, and a few sentences are incomplete (for example, "a lot of the locking that we've seen uh over the last 12 months and yeah"). --- ## 6. Self-Check - **Named roles verified in source:** Speaker name, Tailscale affiliation, and Aperture responsibility are stated in the transcript. No roles were added. - **Promotion flagged:** Yes, with explicit promotional segments and forward-looking statements identified. - **Quality judgments excluded:** This record describes who said what and what sources were named. It does not evaluate the accuracy or merit of any claim. - **Unverified items marked:** Speaker experience claims, roadmap statements, and unsourced industry characterizations are labeled as such. Well, hi folks. Uh, thanks for joining me here for the last talk on this stage uh of the uh AI engineer world's fair this year. So, um, yeah, thank you for joining me. My name is Remy. I am on the strategic projects team at Tailscale. Uh and more specifically I actually work on our AI gateway aperture and today I want to talk a little bit about you know what it means or you know what it's going to take uh to have an AI future without the lockin. I think a lot of the locking that we've seen uh over the last 12 months and yeah so if we take a look real quick at the you know the last 12 months you know I think it can really be you know described quite a bit by uh the phrase token maxing and I I don't mean that in the kind of you know negative sense that you might kind of hear outside of the you know kind of AI enthusiast circle. I mean that in the sort of you know it's been a year of trying to figure out you know what we can do with million token context windows uh with the ability to get context from anywhere right not just you know MCP being everywhere but you know also having coding tools right that are great at finding context using the CLI bringing that context into the model uh bringing it you know bringing it into you know everything that you need and and and going from there right it's it's been pretty cool right I mean we've gotten It's really made truly agentic coding possible having all of this right. So yeah, it's I mean it's it's quite cool and you know we're going to see uh not just agentic coding but you know other agentic use cases continue to take off uh as we as we do this. So um you know token maxing is cool but right it has it has problems right there's obvious problems right it's very expensive right it's it's it's an extremely expensive endeavor uh you know I think the last 12 months are you know have maybe been marked by hey we have you know unlimited AI budget uh and that's changed here in the last few months uh in terms of you know what it what it means to you know what it means to use AI um it also I mean you know it's expensive like just generally if you're doing it, you know, right, but it also it does incentivize what I would call like wasted tokens by default. I mean, I I work on an AI gateway. We use our AI gateway. I get to see all of the usage from the AI gateway. You would not believe the number of sessions that are hundreds or thousands of messages long where nobody clears the context. They just kind of keep going. I mean, it's helpful sometimes, but you know, it's it's it's pretty surprising. you will have you know tens of millions of tokens of a session uh where it's the cash is basically maxed out maxed out the entire time uh it gets really expensive um and the other thing is right so you know this has been an exciting time right people are doing new things building new things uh but this has led to you know in many cases very vertically integrated tool chains uh cloud code is obviously you know a big one codeex right everybody's kind of building up their you know for all the way from the model you know all the way all the way up to the top uh you know building their uh you know building their agents, building their tool chains. So um all right there's a lot of problems there and so you know my question right in terms of this is uh you know how do we address these problems in particular how do we address the lock in and you know how do we actually improve uh how we use AI such that you know we can use it for things that aren't just you know kind of essentially frontier task or where we're just kind of wasting wasting money wasting tokens uh you know that sort of thing. Um, but first I want to talk about how not to address the problem. Uh, and this is a super super common thing. So, uh, over the last 6 months, we've been talking to I've literally been talking to hundreds and hundreds of folks about their AI spend, their AI usage, right? That's the kind of benefit of working on an AI gateway. Folks come to you, they're like, I don't know what's going on. I know we're using a lot of AI. I get, you know, uh, I get an anthropic bill every day. Uh, I get a codeex bill every day. Um uh so you know I it's not just that you know I see it all the time. So you know what do I do? They usually come with one kind of solution in mind uh and that is oh we're just going to like uh everybody seems to be using cloud code right now. So we're just going to like consolidate on a single vendor maybe two vendors right maybe two vendors. So that's uh they're like I need to get a handle on this. How do I get a handle on this? How I get you know oh okay yeah I will say everybody's going to use this right we're just going to use this thing. um that is not not a great idea. uh you know I might seem obvious sitting here it's maybe not a great idea but you know when you're in that panic zone of we've spent a lot of money and we don't know how exactly uh and we you know finance is talking to you it is being like what are you doing uh um you know like we said to you know use AI but maybe not use that much AI uh and you know it it's a pretty natural response to be okay I'm going to consolidate everything down right okay I'm gonna have a control on it I'm gonna have you know I'm gonna have control on this uh control of this Um and you know I think this is kind of peeking at what we're going to see the next 12 months and what people are actually trying to do uh at least at the you know kind of frontier is uh what I would like to call ROI maxing right we're not going to be token maxing we're going to be ROI maxing uh and you know it's it's really just a necessity right it's not uh um and I don't mean that when I say ROI maxing I don't just mean cutting costs right I mean actually proving some sort of value uh to the thing that you're doing uh and not only that but also kind of opening things up, right? So if the only thing you have is like two frontier models and their entire stacked ecosystem, that's a certain cost level where you you it just makes things certain things untenable, right? You can't actually expand AI usage into the company if that's the only way you can use AI. Um and so you know when we think about ROI maxing, we first need to kind of think about like what are the components of an internal AI deployment. I'm not talking about like a hey, we've built this new AI feature in our app. I mean, you know, internal internal to your organization. And so, you know, there's four things, right? There's the LLM or LLMs, right? There's that layer. Uh there's sort of the data connector layer. I mean, this is MCP for a lot of people, but in reality, it's, you know, CLI or APIs or really just any way that you get data could be reading a file, right, into your uh into your agentic system. Uh there's the interface. I think people kind of forget about this sometimes, but you know, it's started with the chat interface. We got two, you know, we kind of got a uh we got, you know, idees with chat. We got, you know, kind of a whiplash back from two into, you know, I'm going to do it on my go to my phone. Uh, you know, to, oh, now the, you know, the hot thing is, you know, claude tag and like I'm going to use, you know, I'm going to use we're just going to interface with agents via Slack, right? So, uh, um, you know, it's another core component and it is a component that you can break out. Uh and then the last one is sandboxes just the environment in which the agent runs right so those are your four kind of uh you know four components really uh and to be ROI maxing uh means to be maxing your you know every opportunity that you have uh for choice to make a choice at one of those layers right you don't have to necessarily have a unique thing at you know oh I've made my bespoke special choice at every one of those layers but it's really important to be able to make a choice at any given layer that you need at any given time. Uh, and it might be different for different teams or different folks, right? It's, you know, it can be pretty different across the organization. I don't totally imagine like an entire engineering team just coding via, you know, slack thread. Might be fun, but I don't totally imagine that, right? Uh, um, but, you know, the entire marketing team might be working entirely uh, you know, in a slack thread, asking data questions, doing other things like that. So it's really important to you know to be able to uh kind of separate these things and right so I I work at tail scale right I work on an AI gateway uh you know maybe no surprise that I'm here to talk a little bit about why an AI gateway uh will help you do that can help you do that and I don't just mean necessarily using our AI gateway I think it's nice but you know just any AI gateway is the um you know is important for this uh so our g AI gateway is uh called aperture and um what it does uh is right it's maybe No surprise here, right? But it let you it sits in between your agents and that interface of that agent uh and you know all of your endpoints uh as well both like LLM endpoints and data you know kind of MCP endpoints API endpoints anything that you have uh and sits in the middle and let you both control and monitor what's happening from a that can be from a like security perspective that can be from a cost perspective that can be from a which we do a lot I want to learn about what somebody is doing or what an agent is doing uh perspective. It's actually really cool when you have logs uh and you can tell your agent to hey uh I it maybe it ran out. It kind of just kind of you know oh it it didn't reach an answer. It just kind of spiraled out of control. Uh maybe happens less now. Definitely used to happen quite a bit. Happens with you know maybe cheaper models more often too. Uh they're usually not great at stopping themselves. Uh but they are great at analyzing where they went wrong after the fact. And so when you have something like an AI gateway, it's really easy to point it back at what you were doing and say, "What happened? Why did this not why did this not, you know, work the way I was expecting it to?" Uh, and so, you know, if you've used an AI gateway, how many people like have an AI gateway, have deployed an AI gateway, use an AI? Yeah. So, if you've used one before, you're probably used to, okay, so I've put all my keys on this gateway. Now, I need to go create synthetic keys for everything that I have, right? So, it's like, okay, I've traded some keys for a bunch of other keys. uh we get to do something a little special. So Aperture is built on tail scale and if you're not familiar with tail scale, tail scale is an identity based mesh network. It's built on wire guard. Uh and what you get to do with tailscale is uh every connection that is made over a tail. So it's a tail scale network uh you know you install the tail net agent everywhere that you need. It can be containers, laptops, GPU servers, doesn't matter. um you install it everywhere and every connection that's now made over that tail net uh comes with the identity of the machine or the user on the other side. So that can be a tag, right? So that can be like hey this is a machine ID. This is a PR review bot, right? Uh and then it would get a separate node ID underneath. This is the instantiation of that PR review bot, right? So you can do that or it can be the user, right? So it's just me using cloud code on my laptop. Uh that can come with it as well. uh in addition to if you're using something like Entra or Octa uh we can actually sync all the groups in uh so you can see the group that that person is associated with. So something on the other side is able to see all of that. We have this really cool library. It's called TSNet. It's an open source library. You can use it to build these sorts of apps. That is what we use to build aperture. That is what is actually happening here. So it is able to see the uh it is able to see the identity of that user or that agent on the other side of the thing on the other side of the connection. And so you know again just reminder components of a you know an internal you know AI deployment uh right so there's the LLM side of things there's the data side of things the interface side of things and the actual sandboxes uh sandbox and environment in which things run uh side of things. So, um, when it comes to using Aperture, uh, you know, as far as LLMs go, this is pretty similar, I think, to, you know, what you might expect from other other AI gateways, except now, you know, here I have a, you know, a configured gateway. Uh, it's got anthropic, open eye, versel, you know, it's got even anthropic oath, uh, you know, bedrock codex, everything, uh, that you can imagine. I put all those keys on that gateway and then now when I want to actually connect to that LLM provider as an agent or a user there's no I don't have there's if I had this expanded there's there's no extra screen for create these API keys create these identities create this once I connect to the once I actually connect to the gateway it knows who I am it knows what I am and it's like oh you have access to this models this provider right you can use GLM 5.2 uh you get this budget with GLM 5.2 too, right? You get unlimited budget with a cheap model and you get an extremely narrow budget with a Frontier model, right? Those sorts of things. Uh there's no extra setting here. It's just add your add your API key and go. So again, we've taken all the LLM layer and we've put it into a you know, put it into one place that you can then, you know, kind of expand out within your organization. Again, doesn't matter if you're trying to switch models or providers, you can do that. Uh the next thing is the data side of things. So, uh, in addition to, you know, being an an LLM gateway, and this is actually why I said AI gateway at the beginning. So, I know that seems kind of nebulous, uh, but I'm actually I very specifically chose that term because we're both an LLM gateway and an MCP gateway and we even do off like, uh, injection for just standard APIs and reexpose those as MCPs, too. So, uh, you know, kind of intentionally choosing the broader term here. Um, but with this, as you can see here, uh, in Aperture as an admin, you can go in and configure whatever MCP servers you want. We have kind of built-in ones where we've, you know, checked them out, make sure they're easy to easy to work with, all those sort of things. If you've ever deployed an MCP or used MCPS within your organization, you'll know that, uh, there's a standard, but everybody follows a different version of that standard. Uh, you know, you can tell exactly when an MCP server was built by exactly how they decided to do off. Uh yeah, it's it's a great uh you know kind of archaeology lesson uh when you go through when you go through setting up MCP servers, but we try to make that a very simple process here, right? We've kind of gone through this. Uh so you as an admin can say, "Hey, I've configured these. These are the ones that are allowed." You can use that same identity information to say, "Hey, the marketing team, you know, they get access to Google Calendar and Gmail, but the engineering team, they'll also get access to GitHub, right? uh you can even say things like oh the engineering you know this side of the engineering team gets access right access to GitHub or readonly access to GitHub you can do all all sorts of things here and what's really nice and again talking about decoupling is all of the authentication happens here so if you have folks that want to use cursor or they want to use cloud code or they want to use codeex or they want to use some special agent framework that they're you know some new hot thing that they've you know they've discovered or built or whatever uh all they have to do is hit that endpoint and if they're on the tail net everything's authenticated as them or as whatever the agent or as whatever you know the tag device is. Uh there's no like oh I have to go reth into everything and then reauth into everything and then reth into this new tool. Uh we make the switching cost right of like actually trying out tools and trying different workflows and different experiences a lot a lot easier with this sort of thing. It's all those little points of friction uh that lead to use cases where like you know people just don't try new use cases. They don't try to uh you know oh maybe I should experiment maybe I can make this cheaper maybe I can make this faster. Um you know next thing talking about the interface so uh with Aperture we actually have a built-in chat UI surprise uh you know everybody seems to have a chat UI we have a built built-in chat UI um however it is built on top of APIs that you can also leverage yourself. Uh so our whole goal with Aperture is to actually provide you with the same sort of you know tools and interfaces and or tools to build interfaces that we can. So whether or not you want to build a Slackbot or you want to build uh chat UI or you want to build you know some new thing that nobody's thought of some mobile app kind of thing you know it's voice only interface right uh you can you can do that. The whole point is you're going to be able to do that with Aperture. We want to support all of those different individual uh use cases. But again you can see it has connectors. Those are all again set up through Aperture itself. They're not set up in the chat UI. You're just getting them uh by way by way of Aperture. Uh and then the last thing, so this is uh you know, so we talked about interfaces, you know, a little bit. It's, you know, it's again, it's very important uh that you be able to kind of use all of that. Uh actually one other thing that's coming very soon here is um uh if you've ever I don't know if anybody's ever tried to set up like open web UI or you know or like Libra like or any of the like kind of like any of the like oh I'm going to set up like a hey this like you know this interface or this interface uh one big thing is you always lose identity information and then suddenly all of the you know everything's coming from the UI right everything's it's like oh everything's just coming from the system uh we're actually going to be adding a full like pass through O uh so you'll be able to actually see all of the like user identities and machine identities if for some reason they're using whatever interface that you have. Uh you should be able to set it up so that you can pass through the O and you still get all of the same visibility that you're kind of expecting from a from an AI gateway as well. Um and then yeah the last thing is right this is actually the I think the area in which um folks are doing pretty good. There's actually a lot of choice already people are trying to do this. Uh I think that's because uh the definition of a sandbox I'm sure I could poll everybody here and would have a different everybody would have a different definition of exactly what a sandbox is uh or what they need from a sandbox or what they want from a sandbox. Uh and so you know there's we're kind of in that let a thousand flowers bloom phase of you know there's a sandbox provider for every sort of whim and need and you know and want out there. Um but uh you know and folks are doing a pretty good job of integrating them but again it's a really important thing. We're actually going to be adding kind of an orchestration layer as well for sandboxes to a gateway. uh again trying to use that term AI gateway very loosely intentionally very loosely because in order to again kind of make the best choices for all of the different endpoints that you're you know environments LLMs data connectors all that sort of thing uh you really do need uh to kind of have a place to sort of see them all uh and then a place that brings them all together such that you can then use any other one so right kind of a almost a universal translation layer in a way um but yeah so that's that's mostly that and you know like I said is a reminder here uh is you know ROI maxing really is all about maximizing choice uh to prevent lock in right so you know even if you don't use Aperture I do highly recommend you start considering an AI gateway you look at an AI gateway it's good just beyond like actual you know oh you know I don't have to you know uh have three different providers configured or switch between them this way you know uh it's good beyond that right it's great for just you know understanding what agents are doing even if it's just you as an individual we have so many people who or just as an individual they have their usage behind an AI gateway just to have an idea as to what their agents are doing because you know if you're running uh very commonly uh is people are running something like openclaw with on tail scale and it's on their tail net and they and it's just running in the background and they're like oh my god it did something I'm not sure what it did it did something it was not supposed to do with something like an AI gateway you can actually go back and you can see and you can look and you can actually diagnose where it went wrong what happened Um so you know it's it's not just lock in right it's actually just experimentation and understanding as a whole uh is you know is very important to use something uh like a like an AI or LLM gateway. So um that's it uh for my talk just short and sweet uh thank you very much for attending I this is the the last talk in here of the day so uh I hope everybody's had a really great conference I know I love the AI engineer world fair uh and I hope you do too. So thank you