Talk by Laurent Gil (co-founder and president, Cast AI) and Žilvinas Urbonas (engineering lead, Kimchi) on measuring cost per task rather than cost per token for AI coding agents, and on an outcome-aware open-source coding harness that selects proprietary or open models per task. The description states the speakers share three months of internal results with 300 employees and demo Ferment (milestone-based autonomous coding runs deploying to staging), Teleport (remote agent sessions that continue after the laptop closes), and Studio (team collaboration with a shared Kanban board). Chapters cover token budgeting, model switching, the harness design, the Ferment and Teleport demos, and Studio. Recorded at the AI Engineer World's Fair 2026 in San Francisco. Claims such as 2.5x savings and 62% of engineers are presented in the description and chapter titles as the speakers' reported results, not independently verified here. # Video Identity Observations ## Source Overview - **Title:** *"Stop Rationing Tokens: Let the Harness Pick the Model — Kimchi by Cast AI"* - **Format:** Conference talk by two company representatives: **Laura, co-founder and president of Cast AI**, and **Zilvinas, engineering lead for the Kimchi coding platform**. ## Self-Description Evidence (Creator Claims) - Company identity and origin: *"My name is Laura. I'm the co-founder and president of Cast AI, who is the inventor of Kimchi."* - Scope claim — full SDLC, not just a harness: *"we are not just providing you the coding harness, we want to provide you the full software development life cycle solution."* - Open source positioning: *"whatever you're going to see, it's open source"*; later qualified: *"The harness is entirely open source. Studio and Teleport, you need a Google account."* - Product suite claims: **Kimchi harness** (model selection), **Ferment** (long-running autonomous tasks with scoring), **Kimchi Teleport** (remote sandboxes), **Kimchi Studio** (team/enterprise collaboration). - Internal dogfooding claims: 300 employees, two-thirds developers; *"we ended up spending too much on cloud code like half a year ago... we started working on our own harness"*; claimed results of **2.5x savings** and **1.5x more tokens** over three months; *"62% of our engineers are only using Teleport."* ## Intended Audience (Observable) - **Engineering managers and enterprise developers** facing AI cost pressure: framing like *"Our jobs as managers is... to ensure... tokens [are] unlimited and inexpensive"* and *"token maxing or the token cost is a thing for the enterprises."* - **Enterprises with compliance needs:** *"We can ship it to your on-premise deployment if you're like a financial technology company."* - **Live conference audience:** direct address (*"All right, guys. Very nice to meet you all"*), booth invitation (*"We are in UG4"*), and references to *"a pizza team, a team of five to 10 guys."* ## Claimed Scope - **Cost optimization via autonomous model selection:** a harness that selects *"the right model for the task at the right time based on the outcome of the task,"* citing cost-per-task disparities (e.g., Gemini 3 at $3.5/M tokens but $705 per task). - **Quality-gated autonomy:** artifacts complete only when scoring is *"at least a B"*; deliberately stops at staging: *"We still believe there's like some little human that needs to be in the loop before going to production."* - **Vision beyond current state:** automated production deploys via observability/SLO checks described as *"things that are coming."* ## Claims vs. Demonstrated Patterns - **Claims:** internal savings figures, adoption percentages, and model-selection shifts (Kimi 2.6 → Minimax 3 on June 12) are asserted with screenshots but not independently verifiable from the transcript. - **Demonstrated:** a product-marketing presentation pattern — problem framing (token costs, Uber/Anthropic anecdote), internal origin story, live product walkthrough, and conference-booth call to action — indicating the channel presents as **vendor thought leadership / product launch content** rather than neutral education. ## Notes for Synthesis - Transcript contains transcription artifacts ("preparatory models," "PyMon OS SDK," "Ziminus"), suggesting an auto-generated transcript; product/name spellings should be verified against other sources before channel-level conclusions. - This single video supports, but does not alone establish, a channel identity around **AI coding-agent cost optimization and enterprise developer tooling**. # Delivery Observations: "Stop Rationing Tokens: Let the Harness Pick the Model" (Kimchi by Cast AI) > **Evidence basis:** Transcript only. Quotations are exact. Bracketed numbers (e.g., [13]) are segment markers from the transcript. The transcript appears auto-generated, with visible errors such as "Zivinas," "Ziminus," and doubled words, so some observations about wording reflect the transcript's text rather than confirmed speech. --- ## 1. Speaker Structure - **Two named presenters** introduce themselves in the transcript: - Laura: "My name is Laura. I'm the co-founder and president of Cast AI..." [13] - Zilvinas: "Hey guys, I'm Zilvinas. I'm leading engineering for a Kimchi coding platform." [36] - **Division of labor is observable:** Laura frames the business problem and the vision. Zilvinas covers the engineering mechanics (harness, Ferment, Teleport, Studio). - **Turn-taking is marked** by `>>` in the transcript and reinforced by explicit handoffs (see Section 5). - **Uncertainty:** A line beginning "Guys, I really like this idea" [773] follows a `>>` marker, but the transcript does not clearly confirm which presenter says it. Attribution for a few mid-section lines is uncertain. --- ## 2. Tone and Register ### 2.1 Conversational, Informal Address - Frequent direct address: "All right, guys." [13], "Guys, you can see this at our booth..." [1027], "Thank you, guys." [1063] - Speech is loosely structured, with fillers such as "Um," "So," "Okay," and "Yep." ### 2.2 Enthusiasm Markers Positive intensifiers recur, mostly from Zilvinas and in later segments: - "That's so cool, huh?" [862] - "This is really cool." [1017] - "It's really good." [1027] - "I'm so excited to be here with you." [48] - "Yep." is used as a short acknowledgment between speakers, for example [1017] onward. ### 2.3 Hedged Claims Hedging appears in several places, often softening assertions about future or uncertain points: - "probably most of you know this" [663] - "Also as human, probably we can find and understand and test..." [394–405] - "I think it was not the Wi-Fi. It was the battery that broke." [793] - "We still believe there's like some little human that needs to be in the loop..." [620] **Observation:** Hedges cluster around *opinion or future-state claims*. Factual figures (costs, percentages) are stated with little qualification. --- ## 3. Vocabulary and Recurring Phrases ### 3.1 "Our job" / "Our responsibility" Framing A recurring structure frames the company's role as a moral or operational duty: - "Our job is not to prevent the developer to use coding agent." [136] - "This is our job as managers." [326–337] - "It is our responsibility to do it." [431] Note: The phrase "Our job" appears at least three times, making it the most repeated framing device in the transcript. ### 3.2 Contrastive Pairings Statements are often built on *this, not that*: - "We at Kimchi think this is wrong. Our jobs as manager is to do the opposite." [113–124] - "They don't cost the same, but also not the same per task." [245–258] - "Not the Wi-Fi. It was the battery that broke." [793–804] ### 3.3 Rhetorical Questions and "Imagine" Prompts - "Wouldn't it be nice that you have an automated fashion, automated engine, automated harness..." [258–270] - "Do you think we look at this as human? Absolutely not." [405–411] - "imagine like product manager doing the coding or my dear colleague Laura starting to do some coding..." [675–685] - "Think about it." [981] ### 3.4 Evaluative Closers - "How simple that is." [827–839] - "And yet our favorite tool is this one." [839–851] - "A yet yet a very simple innovation that is so essential..." [972–982] ### 3.5 Comparison Language - "otherwise you cannot compare apple with apple." [201–213] - "Same outcome." [313–326] --- ## 4. Analogies and Concrete Illustrations | Concept | Analogy or example used | Location | |---|---|---| | Token limits | A laptop with a one-hour battery that can only be charged once a day | [102–113] | | Lost connectivity | A developer coding on a plane, interrupted | [773–804] | | Cross-model comparison | Cost per token versus cost per task | [191–245] | | Remote work continuity | Coding continues after the laptop is closed, on holiday, or in transit | [839–862] | | Team collaboration | Kanban-style review board for a team of five to ten | [961–1006] | **Observation:** The laptop-battery analogy is the only explicit metaphor for the core problem. Other explanations rely on concrete numbers and product walkthroughs. --- ## 5. Teaching and Presentation Habits ### 5.1 Stated Roadmaps Presenters preview content before delivering it: - "I'm going to show you what we built, how we built it, and the result of it." [158–169] - "And then Zivinas will show you a few more things..." [169] - "I'm going to show you the evolution of the selection of the model over time..." [347–359] ### 5.2 Explicit Handoffs - "Ziminus is going to tell you a little bit more of how it works" [431–442] - "Go, Ziminus." [442] - "Okay, so one more thing to mention." [442–453] - "Yep. So, there is more to that." [862] ### 5.3 Reveal-Then-Explain Structure Figures appear first, then interpretation follows: - A chart is introduced, then "This is a graph of what would have been our cloud bill by day." [292–302] - Model comparisons: "look at Gemini 3 Plus..." followed by cost detail [213–224] ### 5.4 Pointing to Visuals - "look at this" [302–313] - "Look at the this This is a screenshot of Teleport." [762–773] - "Look at the" (transcript fragment) [762] ### 5.5 Self-Correction in Live Speech Corrections appear inline without apology markers: - "At Kimchi, we have 300 developers. Oh, sorry. 300 employees." [136–158] - "the task to complete with Gemini 3 was $305. Sorry, $705." [224–234] ### 5.6 Audience Invitation - "Guys, you can see this at our booth if you want to uh to see more of it." [1027–1038] - "So, come see us. We'll be delighted to speak to you." [1050–1063] --- ## 6. Numeric Delivery Figures are presented frequently and with minimal framing: - "$500 million on Anthropic alone in one month" [48–59] - "2.5 times is the amount of savings" [292] - "The cost per token is 1.5 and the task the cost for the task is $148." [234–245] - "62% of our engineers are only using Teleport when they are coding now." [839–851] - "every 2 hours, every 3 hours" [497] **Observation:** Numbers are often paired with a *comparison* (cost per token versus cost per task) or a *time reference*. Several figures are stated in fragments that may reflect transcription errors (e.g., "The cost per token is 1.5" lacks a unit). --- ## 7. Disfluency and Repetition (Transcript Artifacts) Repeated words appear in several places. Some may reflect delivery, others transcription: - "once once once a day" [102–113] - "You see what what I like" [573] - "a a a" constructions, e.g., "it's a a 5-minute install" [1038–1050] - "A yet yet a very simple innovation" [972] - "Kimi Kimi 2.6" [394] **Uncertainty:** Whether these are deliberate emphasis or transcription duplication cannot be determined from the text alone. --- ## 8. Pacing and Structure - The talk moves from **problem** (token cost, [48–90]) to **thesis** ([113–136]) to **evidence** (charts and model data, [158–431]) to **product demonstrations** (Ferment, Teleport, Studio, [442–1038]). - Transitional markers signal section changes: "Moving on." [697], "So, there is more to that." [862], "And this is yet another thing we built for ourself." [928]. - The closing is brief and promotional: a booth location, an install note, and a thank-you [1027–1063]. --- ## 9. Summary of Stable vs. Isolated Patterns ### Stable-looking (repeated within transcript) - "Our job" / "our responsibility" framing (3+ instances) - "Yep." as a conversational acknowledgment - Explicit roadmaps and handoffs - Enthusiasm intensifiers ("really cool," "so cool") - Rhetorical questions followed by immediate answers ("Absolutely not.") ### Isolated or uncertain - "Moving on." appears once. - The laptop-battery analogy appears once. - "Imagine like..." appears once or twice; not enough to establish a habit. - Speaker attribution for some mid-transcript lines is unclear. - Repeated-word artifacts may be transcription noise rather than delivery style. --- ## 10. Limitations - The transcript is a single recording with no audio, visual, or prosodic information. Tone is inferred from wording only. - No curated metadata was supplied, so delivery context beyond the transcript cannot be confirmed. - Observations about hedging and enthusiasm reflect word choice, not vocal delivery. # Video Observation: "Stop Rationing Tokens: Let the Harness Pick the Model — Kimchi by Cast AI" ## Speakers - **Laura** — Co-founder and president of Cast AI, identified as "the inventor of Kimchi" (self-described in transcript). - **Zilvinas** — Leads engineering for the Kimchi coding platform. ## Context and Motivation ### Claim: Token costs are out of control for enterprises — Laura Cites two recent examples as evidence: - A company in India that **spent $500 million on Anthropic alone in one month**. - A "famous tweet" from a person Laura calls "the city of Uber" claiming that in the last four months they used Anthropic's **entire annual budget**. ### Claim: Rationing tokens is the wrong managerial response — Laura - Companies have responded to cost pressure by limiting developers' token usage. - Analogy: limiting tokens for a developer feels like *"you can use your laptop, you have an hour battery and you can only charge once a day."* - Stated position: *"Our job is not to prevent the developer to use coding agent. Our job is to make sure they can use it as much as they want for as long as they want."* ### Corroborating claim — Zilvinas - Cast AI itself was *"spending too much on cloud code like half a year ago"*, saw cloud bills "skyrocketing," and built the harness for their own use before open-sourcing it. ## Core Thesis: Cost Per Task ≠ Cost Per Token ### Claim: Token price is a misleading cost metric — Laura - Cites an (unnamed, described as "famous") university white paper comparing **cost per token vs. cost per task at the same quality**. - Examples given from the study: - **Gemini 3 Plus**: appears inexpensive at ~$3.5 per million tokens, but cost **$705** per task (Laura self-corrects mid-sentence from "$305" to "$705"). - **Minimax 2.7** (described as "one of my favorite"): $1.5 per token, **$148** per task. - Conclusion: *"the models are not the same. They don't cost the same, but also not the same per task."* ### Claim: Automated model selection by a harness solves this — Laura - Proposes an "automated harness" that selects the right model per task, at the right time, **based on the outcome** of the task. ## Reported Results (3 months of internal use) All figures presented by Laura, for Cast AI's ~300 employees (2/3 developers): - **2.5× savings** compared to what their Claude/cloud bill would have been. - Token usage **increased 1.5×** while cost to the cloud provider **decreased 1.5×** — "same outcome." - The harness shifted model selections autonomously: Kimi 2.6 dominated early June; **Minimax 3 dominated after June 12**. - Key rhetorical point: *"Do you think we look at this as human? Absolutely not."* — humans wouldn't catch and migrate to a new model; an automated harness "obsessed with token cost" will. ## Platform Components (Zilvinas, with Laura's commentary) ### Ferment (long-running autonomous tasks) - Built on the PyMon OS SDK (as stated); contributed back to open source. - Runs coding tasks autonomously for **2+ hours**; human-in-the-loop involvement estimated at roughly every 2–3 hours. - Breaks tasks into **milestones**, produces a **score** at cycle end. - Artifacts are considered complete only with a **score of at least a B**; users can ask the agent to iterate toward an A. - Runs a full build/fix loop: build → check → remediate → rebuild; deploys to **staging** when clean. ### Qualification: staging-only by design — Zilvinas - Explicit limitation: production deployment is deliberately excluded — *"We still believe there's like some little human that needs to be in the loop before going to the production at least at this point."* - Acknowledged future direction: automatic production shipping gated on observability data, SLOs, Kubernetes pod health, and health checks. **This is stated as not yet built.** ### Kimchi Teleport (remote sandboxed execution) - Origin story (Laura): a colleague coding on a plane lost the session — Laura corrects Zilvinas: *"I think it was not the Wi-Fi. It was the battery that broke."* - Synchronizes the local environment to a container running on a hyperscaler (currently Google); laptop can close while work continues. - Claimed adoption: **62% of Cast AI engineers use Teleport** for coding as of Laura's last check. - Available as SaaS or on-premise (for e.g. financial technology customers). ### Kimchi Studio (team collaboration layer) - Described by Zilvinas as "a Teleport for teams and enterprises": a Kanban-style board for sharing agent sessions, plans, and work-in-progress. - Rationale: **reading code is no longer enough** — agentic engineering produces change sets too large to review manually (example: a 2,000-line PR from a non-technical colleague), and reviewers must check *intent and specifications*, not just diffs. - Workflow: sessions move between backlog → in progress → review; teammates can claim and answer review questions; work continues when laptops close. - Laura frames it as a simple innovation born of internal need: *"We found it so cool, we decided to make it a product."* ## Availability and Licensing Qualification - **The harness is entirely open source** (Laura). - **Studio and Teleport are not fully open source** — they require a Google account; Laura describes them as "a little bit more complicated" but a ~5-minute install. This is an explicit qualification on the "open source" framing repeated throughout the talk. ## Unresolved Tensions / Contradictions Preserved - Laura's thesis is "unlimited, inexpensive tokens," yet the mechanism described is cost optimization via cheaper per-task models — token rationing is rejected, but cost control remains the explicit goal ("obsessed with token cost"). - The claim of full autonomy (harness picks models, ferment runs 2+ hours) coexists with acknowledged human gates: human review before production, human-in-the-loop every 2–3 hours, and scoring thresholds that can be overridden. - The open-source claim is qualified: only the harness is fully open source; Teleport and Studio require a Google account and are not described as open source. - Minor factual inconsistencies appear in the transcript itself (e.g., "$305… sorry, $705"; "300 developers. Oh, sorry. 300 employees"; Wi-Fi vs. battery origin story), preserved as spoken. - Model names and study figures are presented without citation details beyond "a famous white paper in a university." # Video Attribution Record: Kimchi by Cast AI **Video ID:** 01a15ca7-20f9-7b07-8347-5af5aca9943b **Title:** Stop Rationing Tokens: Let the Harness Pick the Model — Kimchi by Cast AI **Basis:** Transcript and title only. No curated metadata was supplied. --- ## 1. Speakers and Roles ### Identified Speakers - **Laura** - Self-identified as "co-founder and president of Cast AI." - Introduces Cast AI as "the inventor of Kimchi." - Presents the cost, token-optimization, and selection-over-time sections. - **Zilvinas** - Self-identified as "leading engineering for a Kimchi coding platform." - Introduced by Laura as running "the entire engineering team for Kimchi." - Presents the ferment, Teleport, and Studio sections. ### Unattributed Speaker Turns The transcript contains speaker-change markers (`>>`) without names for several turns: - The turn beginning "Guys, I really like this idea" (Teleport discussion). - The turn beginning "And this is yet another thing we built for ourself" (Studio discussion), which may be Zilvinas or another speaker; attribution is not explicit. - Brief responses ("Yep," "It still works," "This is really cool," "Thank you, guys"). **Status:** Speaker identity for these turns is **not stated in the transcript**. Do not attribute them to Laura or Zilvinas without further evidence. --- ## 2. Referenced Sources and Third-Party Claims ### News Items (Unnamed Source) - A company in India allegedly spent **$500 million on Anthropic alone in one month.** - A "famous tweet from Uber" allegedly stated that the company used its **entire annual Anthropic budget in four months.** - *Note:* The transcript reads "the city of Uber," which is likely a transcription error. Source details are not provided. ### University White Paper (Unnamed) - Cited as "the famous white paper in a university" comparing LLM cost per token and cost per task on the same task and quality level. - **Claims reported by the speaker:** - Gemini 3 is "$3.5 per million token" (described as a blended average at study time). - Gemini 3 task cost is **$705** (spoken as "$305," then corrected to "$705"). - Minimax 2.7 cost is **1.5 per token** (unit not stated) and **$148 per task**. - No study title, authors, date, or URL is given. ### Named Vendors and Models Referenced - **Anthropic:** Named as a token-spend provider in the news items. - **Google / Gemini 3:** Named as a model; the speaker refers to "our friends at Google." - **Minimax 2.7 and Minimax 3; Kimi 2.6:** Named as models in the harness selection data. - **"Cloud code":** Referenced as a prior tool the team used, spelled as transcribed. The speaker says the team "ended up spending too much on cloud code like half a year ago." - **"PyMon OS SDK":** Referenced as the basis for the harness. This is a transcription that may not be accurate; the underlying SDK is **not confirmed**. ### Internal Claims (Self-Reported by Presenters) - Cast AI has **300 employees**, about two-thirds developers. - The team has used the Kimchi coding agent for **three months**. - Reported savings: **2.5 times** lower cloud-equivalent cost over the period. - Token usage reportedly increased **1.5 times** while cloud-equivalent cost decreased. - **62%** of engineers reportedly use Teleport "when they are coding now." - Model selection by the harness reportedly shifted on **June 12**, from one model to another, with dates given for June 3 and June 21. - The Kimchi harness is described as "entirely open source"; Studio and Teleport are described as "most of it" open source, with a Google account required for Studio/Teleport. **Status:** All internal figures are presenter-reported. The transcript provides no supporting data, dashboards, or independent verification. --- ## 3. Promotional Segments **The entire video is promotional.** It is a product presentation for Kimchi and its associated tools (Kimchi Coding, ferment, Kimchi Teleport, Kimchi Studio) by the company that makes them. Explicit promotional elements include: - **Calls to action:** "come see us," "guys, you can see this at our booth," "UG4," and "a 5-minute install." - **Product demonstrations and screenshots:** Teleport and Studio are described with on-screen references. - **Product positioning:** Claims that Kimchi is the "inventor" of the product category and that the platform is a "full software development life cycle solution." - **Commercial framing:** Statements such as "we decided to make it a product" and references to on-premise deployment for customers. --- ## 4. Availability and Limits - **Comment availability:** Not provided in the supplied material. - **Curated metadata:** Only the title is supplied. No channel, publication date, description, or speaker bios are included. - **Transcript reliability:** The transcript contains apparent transcription errors, including "Zivinas" vs. "Zilvinas," "Kimi Kimi 2.6," "PyMon OS SDK," "cloud code," "ferment," and "Preparatory models for edge case." - **Unverifiable evidence channels:** - Cited white paper and news items: no links or identifiers. - Internal usage and cost metrics: referenced graphs and screenshots are not included in the transcript. - Unattributed speaker turns: identity unknown. --- ## 5. Provenance Summary | Category | Source | Verifiability in Transcript | |---|---|---| | Industry news (India spend, Uber tweet) | Unnamed | None provided | | Cost-per-task comparison | Unnamed university white paper | None provided | | Model names and selection shifts | Presenter-reported | None provided | | Usage, savings, employee metrics | Cast AI internal, presenter-reported | None provided | | Product features and open-source status | Presenter-reported | None provided | All right, guys. Very nice to meet you all. My name is Laura. I'm the co-founder and president of Cast AI, who is the inventor of Kimchi, which I'm going to show you in a few seconds. Um, the talk is going to be about preparatory models for edge case, but use open source for the rest. And you're going to see some great we have with that with the coding agent called Kimchi. I have the pleasure to be here with Zilvinas. \>\> Hey guys, I'm Zilvinas. I'm leading engineering for a Kimchi coding platform. \>\> So, you have Zilvinas who run the entire engineering team for Kimchi. I'm so excited to be here with you. All right, let's go. So, um, you have you may have seen this in the news. Uh, two two piece of information that came up in the last few weeks. The first one is there's a company in India that spent $500 million on Anthropic alone in one month. Which is again of a good example of tokens that are completely out of control. The other example is a famous tweet from Uber, the city of Uber, who say that in the last four months he actually used the entire budget of Anthropic for the entire year. And this is how much the token, um, the token maxing or the token cost is a thing for the enterprises. And you're going to see how we fix this. Companies responded this by limiting the use of token. You may have heard this many, many times with our friends developers. And as a developer, when you limit the amount of token, it feels like this. It's like, "Hey, you can use your laptop. You have an hour battery and you can only charge once once once a day." That's how it feels to limit the amount of token for developer. We at Kimchi think this is wrong. Our jobs as manager is to do the opposite. It's to ensure that we want to make the tokens unlimited. And inexpensive. Our job is not to prevent the developer to use coding agent. Our job is to make sure they can use it as much as they want for as long as they want in a completely unlimited fashion. This is why we started Kimchi. At Kimchi, we have 300 developers. Oh, sorry. 300 employees. 2/3 of those are developers. The Zivinas run that. And we've been using this coding agent for the last 3 months. I'm going to show you what we built, how we built it, and the result of it. And then Zivinas will show you a few more things that make the coding agent really, really cool. It's coming now. We built an entire token optimization based on this idea. What you see on screen is the true cost of LLM model for the same task. It was made by the famous white paper in a university. And here's what they say. On one side, you see the cost per token. On the other side, you see the cost per task for the same task with the same quality. That's is the important, otherwise you cannot compare apple with apple. And here's what is written here. First, they are all different. But also, look at Gemini 3 Plus. Our our friends at Google. Well, it seems to be a relatively inexpensive model at $3.5 per million token. This is a blended average at the time this study was made. But actually the task to complete with Gemini 3 was $305. Sorry, $705. And look at the others. Minimax 2.7 is one of my favorite. The cost per token is 1.5 and the task the the cost for the task is $148. It means the models are not the same. They don't cost the same, but also not the same per task. And that's why we built our coding agent. Wouldn't it be nice that you have an automated fashion, automated engine, automated harness that will select the right model for the task at the right time based on the outcome of the task? That is what we did with Kimchi Coding. That is what Zilvinas did with his team. And here's the result. We've been using it for 3 months. 2.5 times is the amount of savings we have over cloud for the period. This is a graph of what would have been our cloud bill by day. The purple one on top is what would have been cloud cost over time. The blue line is what we paid. Same outcome. But look at this. The number of tokens increased by 1.5 times. Okay, we all experienced that. The cost to cloud decreased by 1.5 times. This is our job as managers. We provide to our team a coding agent that is essentially unlimited because of that coding agent is growing 3x less, 2.5 times less than the gross of token for the task. There's another result. The multi-model I showed you before. I'm going to show you the evolution of the selection of the model over time by the harness. And here it looks like you see there is a lot of orange early June and there is a lot of yellow late June. There's a complete shift, meaning the harness is selecting a very different model. It happened the 12th of June. And this is what happened the 12th of June. The 3rd of June, Kimi Kimi 2.6 is winning by far the most popular of the harness. 21st of June, the one winning is Minimax 3. Also as human, probably we can find and understand and test and be comfortable with a new model coming in. Do you think we look at this as human? Absolutely not. But an autonomous harness, an automated coding agent that is obsessed with token cost is going to do just that. That is what I mean when I say our responsibility is to think about providing our developers with a right cost per token for the same outcome. It is our responsibility to do it. Ziminus is going to tell you a little bit more of how it works because whatever you're going to see, it's open source. Go, Ziminus. \>\> Okay, so one more thing to mention. So, while I mentioned that token maximum is no longer an option for everyone and uh as a manager at the company as the of you guys probably heard, the having unlimited amount of tokens to spend is no longer looking to be like a case and we had to build this harness for ourselves. So, meaning like we ended up spending too much on cloud code like half a year ago. We saw this cloud bills skyrocketing and we started working on our own harness. Behind this is using like PyMon OS SDK, but the goal of that is like to take the open source experience, still contribute back to open source by providing our own harness to open source. Building constructs like ferment which fits into the kimchi theme which is meant for long running tasks where on average we estimate the human in the loop for these tasks is uh typically happening only like every 2 hours, every 3 hours and etc. So, the goal is uh use kimchi harness, you run a ferment process, it asks you a bunch of questions and autonomously runs the coding task for like more than 2 hours. The goal of that is that because it breaks it into milestones, it knows how to implement it and at the end of the cycle you would get scoring for your output so you would actually know what's happening. And going back to previously discussed like the amount of models used with uh kimchi harness, the scoring is there to prove you that whatever model we will pick for you, you will always have like good enough scoring for the output on the task you're implementing. The way whole kimchi platform works, the goal is like uh we are not just providing you the coding harness, we want to provide you the full software development life cycle solution. So, you do a change, it runs a build for you, it checks if something broke, if something broke, it will go back, do the change again, rebuild again and that's how this fermentation process looks. If everything is complete, there is no extra remediation needs to be done, the changes will be deployed to the staging environment. \>\> You see what what I like with the way you guys explain it is the harness select the model, ferment is constantly checking the quality of the code and going back, meaning there is a a higher order model that is checking the quality and thinking that quality is not good enough, please do it again. This is what I was saying earlier by it looks at the outcome. And it's optimizing the token based on the quality of the outcome. That is what is super cool with ferment. \>\> Yep. And artifact is only considered complete once like the output of the scoring is like at least a B as a score. And even for that, like if you don't like the score, you can ask agent to go back, fix it up and like it will try to achieve a A as a scoring. The intention of stopping at the staging is currently by design. We still believe there's like some little human that needs to be in the loop before going to the production at least at this point. We all know that the ultimate future is to avoid that. You go to production, you look at the metrics, you look at the observability data, the SLO metrics you have in your environment. If you're running on Kubernetes, you want to check if the pods are not crashing, if application health checks are passing. So, all of these things need to be done before it ship it ship to production automatically. So, this is like one of the things that are coming. So, the current limitation we stop at staging, but they'll go to production and and further on that. What we believe is that reading code is not enough anymore. So, like probably most of you know this like across this year like the more live engineering or agentic engineering we do in our coding environments, the more code changes we get in our change sets and it kind of becomes unbearable to read, especially as you keep bringing like a little bit of a less technical personnel into your pipelines. So, imagine like product manager doing the coding or my dear colleague Laura starting to do some coding and suddenly he comes up with a PR which is like 2,000 lines long and you need to actually review not just the differences but the actual intent which he wanted to achieve and the specifications which he provided to the coding agent. So, yeah. Moving on. One thing we built more on top of that is the thing we called Kimchi Teleport. So, Kimchi Teleport allows you to close your laptop. It spins up sandbox in the remote environment. It's a secure sandbox where you don't need to care about anything. We can do that both on our SaaS platform. We can ship it to your on-premise deployment if you're like a financial technology company and things like that. So, we will manage these agent sessions for you. We will make sure they are running. They are accessible both from your laptop, from mobile devices. They are running in the background. So, it goes back and ties to the story that I mentioned about the human in the loop is only appearing like every 2 or 3 hours. So, Teleport emphasizes that. You run your process. You teleport it to the remote environment. The young spins up in the background. You don't need to care about it. We will make sure it's the right size, it's the right security requirements and everything else and that will keep on running until the task is done or until the human in the loop needs to be involved, get notifications for some clarification of specs and things like that. \>\> Guys, I really like this idea. Look at the this This is a screenshot of Teleport. I am puzzled that sometimes innovations are always the simplest one. Like this was born out of the idea that one of us was in the plane coding and then Wi-Fi breaks. And what do you do when Wi-Fi breaks? It breaks the coding because it lives inside your thing. I think it was not the Wi-Fi. It was the battery that broke. Not the Wi-Fi. So, you lost So, he That's it. The guy was stopped there. What Teleport does it it takes all your environment, it synchronizes with a container that live inside one of the hyperscaler. Now we can use Google. And locally on your laptop, it feels exactly the same. It just doesn't run inside your laptop. It feels like it. It runs somewhere else in your environment in a container in a Kubernetes cluster inside the place at Google that will continue running even if you close your laptop. How simple that is. And yet our favorite tool is this one. I think the the last one I I I looked at it, 62% of our engineers are only using Teleport when they are coding now. For that problem, they go home and continues. They're in transit, it continues. They're on holiday, they don't look at it anymore. It continues. That's so cool, huh? \>\> Yep. So, there is more to that. So, we built another tool called Kimchi Studio. If that Teleport is something typically individual uses, Studio is kind of like a Teleport for teams and enterprises. The goal behind the Studio is that you have a camera like board on your environment. It again runs in the Teleport sandboxes in the secure environment. The thing here is like you are supposed to share your tasks on work in progress and plans which are made by the Kimchi code harness with open source or proprietary models with your teams. So, for example, imagine you as an engineer you are like working on a task. You started implementing asking for a plan. The plan can be easily reviewed by your colleagues like by your peer engineer, by your product manager or someone else who cares what's happening and this is how you kind of work with your whole team on collaborating on this agentic engineering journey instead of like saying, "Oh, I'm going to type this prompt. I'm going to try to execute it to the production." Only later on to learn that the specification and requirements were incorrect, uh the output is not what was actually wanted by the task and things like that. \>\> And this is yet another thing we built for ourself. We found it so cool, we decided to make it a product. How do you work with your colleague if what you have is a is a is a terminal, a CLI? How do you share that with the others? Well, first he has to live somewhere else. So, Teleport is the base of what we call studio. And studio is a very nice way to visualize all the sessions that are running, including those that are asking for a review. And then someone in the teams can say, "I take it." And I answer the I I answer the question that is for review. Then you answer it, it goes back to in progress until the next one goes into review. And you can have many, many of those in a team in a pizza team, a team of five to 10 guys. Think about it. A yet yet a very simple innovation that is so essential in how our engineers are coding. They can now share in a very simple way, right? It's a It's a browser. It's graphical. They can share and they can take over the task according to the Kanban style of what is in review, what is in progress. They can start new ones with backlog. And you know what? If they close the laptop, \>\> It still works. \>\> It's also cool. \>\> Yep. \>\> This is really cool. \>\> Thank you. \>\> And \>\> Yeah. We forgot to show the screenshot. \>\> This is the one that we took over, right? That's the one. Guys, you can see this at our booth if you want to uh to see more of it. We are in UG4, it's exactly this way across a few streets. But, have a look at it. It's really good. It is open source, most of it, at least. The harness is entirely open source. Studio and Teleport, you need a Google account. So, it's a little bit more more complicated, but it's a a 5-minute install. And then you get your Teleport set up, as many sessions as you want inside your Google account. And then you can share with share it with your colleagues with Kimchi Studio. So, come see us. We'll be delighted to speak to you. Thank you, guys. \>\> Thank you, guys.