Talk by Mardu Swanepoel of Flinn AI identifying four shared patterns among leading AI agents (Harvey, Cursor, Manus, and Claude): focus modes that constrain the action space to improve output quality, transparent execution that surfaces tool calls and reasoning to build trust, personalization that optimizes for speed to understanding rather than just speed to output, and reversibility that bounds the downside of mistakes. Includes concrete examples of how each company applies these patterns, such as Cursor's line/file/conversation-level rollbacks and parallel model runs, Harvey's firm-specific legal playbooks, and Claude's live task list with visible tool call inputs and outputs. # Video Identity Observations — "What the Best Agents Share — Mardu Swanepoel, Flinn AI" ## Intended Audience - **AI/agent engineers and builders**: The speaker addresses people who build agents — "what we can learn from them in order to ourselves actually build agents in a much better way" — and repeatedly frames benefits "for us as engineers." - **Practitioners familiar with the AI tooling landscape**: References to Cursor, Claude, Manifold, and Harvey are made without introduction, assuming audience familiarity. - Likely a **conference or meetup talk** format: opening "before I jump into sharing with you," closing "thanks a lot... Please reach out if there's more questions," with applause noted in the transcript. ## Claimed Scope - **Analysis of existing AI agent products**: "I'm going to have a look at four of what I believe potentially to be some of the best agents that we have access to at the moment." - **Four named patterns**: **focus modes** (constrained action/input space, e.g., Cursor's planning/debug modes), **transparent execution** (e.g., Claude co-work's tool-call visibility, Manifold's task progress), **personalization** (e.g., Harvey's playbooks and memory, Claude skills), and **reversibility** (e.g., Cursor's granular rollback, Harvey's Word add-in). - Structure is explicitly stated: for each pattern, "what exactly this pattern entails... the value that it adds to you using them, and then thirdly, show you quickly how does that actually look in real life." ## Explicit Self-Description - **Speaker's profession**: "a quote that I actually keep quite close to me when I personally develop um agents" — the speaker personally builds agents. - **Attribution via title**: "Mardu Swanepoel, Flinn AI," suggesting affiliation with Flinn AI (no further detail given in the transcript). - **Opinion framing**: Positions claims as personal judgment — "what I believe potentially to be some of the best agents," "in my opinion, a really, really powerful way." ## Observable Presentation Patterns - **Evidence-driven, product-example style**: Every pattern is grounded in named products with concrete UI/behavior descriptions (e.g., "top right, it has like a progress list or a to-do list"). - **Builder-to-builder teaching tone**: Practical framing like "Let's drop a bunch of tools. Let's really refine our system prompt. Let's optimize our e-vals." - **Intellectual framing device**: Opens with the Picasso "steal" quote to justify deep study of existing agents before building. - **Concise time-boxed delivery**: "quite a lot for a short amount of time" — dense, fast-paced talk. ## Notes on Limits - No channel-level metadata (subscriber counts, channel description, series naming) was supplied; identity observations rest solely on title and transcript. - Speaker expertise is demonstrated but no formal credentials, employer details beyond the title attribution, or channel mission are stated. # Delivery Observations — Mardu Swanepoel, "What the Best Agents Share" ## Tone - **Conversational and low-key**: The talk reads as informal and unhurried, with hedged claims rather than bold assertions: *four of what I believe **potentially** to be some of the best agents*. - **Self-positioned as a learner, not an authority**: *"Um so, in my opinion, a really, really powerful way in which Cursor is using modes"* — evaluations are framed as personal opinion. ## Verbal Fillers - Frequent **"um"** throughout nearly every paragraph, e.g., *"Um but instead, he referred to going and looking at something"*; *"Um additionally, it also enables the user…"* - Occasional false starts and self-corrections: *"what are the potential issues with your with your code"*; *"roll back or or choose"*; *"and and carry on"*. ## Structural Habits - **Explicit roadmap at the start**: *"I'm going to have a look at four specific patterns that these agents use"* — then a per-pattern template: *"I'm briefly going to touch on what exactly this pattern entails, importantly what is the value that it adds… and then thirdly, show you quickly how does that actually look in real life."* - **Numbered signposting**: *"The first one is what I call focus modes"*; *"The second pattern is transparent execution"*; *"The third pattern is personalization"*; *"And the last one is then reversibility."* - **Rhetorical question-and-answer habit**: *"Um what do we get from this? Well, first of all…"*; *"So, we go into a planning mode… Um what do we get from this?"* — used to transition from definition to value. ## Recurring Vocabulary & Phrases - **"Really" as an intensifier**, heavily repeated: *"really trying to make"*, *"really use this process"*, *"does this really well"*, *"really, really powerful"*, *"really, really well"*, *"really critical"*. - **"So" as a connective** opening many explanatory clauses: *"So, we go into a planning mode"*, *"So, by going into a specific mode"*, *"So, if we know what the worst-case outcome is"*. - **"Um and this is really where…"** as a pattern-definition opener: *"personalization. Um and this is really where we try and give the agent…"*; *"reversibility is really the ability…"*. - **"What we're trying to…"** framing for pattern purposes: *"what we're trying to do in this instance is really trying to make…"*, *"what we're trying to get here is to optimize…"*. ## Teaching Habits - **Concept-first, then concrete product example**: Each pattern is defined abstractly, then grounded via Cursor, Claude, Manifold, or Harvey. Cursor is the go-to example, appearing in **three of four patterns** (focus modes, reversibility, plus debug/planning detail). - **Quotation to open**: begins with a Picasso quote and spends time unpacking its meaning before the core content — an attention-framing device. - **Value-first justification**: repeatedly asks what the benefit is before showing the mechanism: *"first of all, the biggest benefit is for us as engineers"*. - **Second-person and first-person framing**: mixes "we/us as engineers" with direct address — *"I give you a task and you come back with just simply the results"*, *"makes it easy for you to not have much downside"*. - **Self-disclosed expertise limits**: *"I'm not a legal expert, but as I understand…"* — flags uncertainty explicitly when outside domain knowledge. ## Communication Patterns - **Analogy to human collaboration**: *"shift from uh delegation to collaboration"*, describing agent trust as analogous to a colleague sharing their process. - **Memorable distilled phrasing**: e.g., *"doing the right thing and not just something"* — stated twice, suggesting a deliberate catchphrase. - **Brief, rushed close**: *"That was I think quite a lot for a short amount of time. I hope it was useful. Please reach out if there's more questions."* — an abrupt wrap-up contrasting with the structured body. ## Uncertainty Notes - Single talk, so patterns like the Picasso-quote opener or the exact "firstly/secondly/thirdly" template cannot be confirmed as stable habits across talks. - Audience reaction is minimal in the transcript (one snort, applause at the end), so audience-engagement patterns cannot be reliably assessed. # Observation Record: What the Best Agents Share — Mardu Swanepoel, Flinn AI ## Framing & Approach - Speaker opens with a **Picasso quote on "stealing"** — he clarifies Picasso did *not* mean taking something and presenting it as your own, but **studying deeply, understanding, making it your own**, then producing something better/unique. - His stated goal: examine **four agents he believes are "potentially some of the best"** and extract four patterns so attendees can build better agents. (Note: he names patterns but only cites examples — Cursor, Claude co-work, Manifold, Harvey, Claude.) - For each pattern he promises: (1) what it entails, (2) its value, (3) real-life examples. ## Pattern 1: Focus Modes - **Definition:** Constrain the agent's action and input space (e.g., planning mode, research mode). - **Value — engineer perspective:** improve output quality on a smaller constrained space; drop tools, refine system prompt, optimize evals on the small space first. - **Value — user perspective:** in "do anything" agent UIs, users don't know how to get the best result and have inflated expectations; modes **align expectations and tailor input/behavior**. - **Example — Cursor:** mode switcher via dropdown; **planning mode** writes no code, only plans and asks questions ("you should be fine with it because that's what you signed up for"); **debug mode** uses a hypothesis-driven approach, spinning up a dedicated debug server and pushing logs there. - His assessment: "a really, really powerful way" Cursor uses modes. ## Pattern 2: Transparent Execution - **Definition:** Make what the agent is doing, using, and thinking extremely clear; shift from **delegation to collaboration**. - **Value:** - Builds **trust in output** — results alone are trusted less than results plus process (thoughts, reads, assumptions, uncertainties). - Enables **early user intervention**, reducing waste (e.g., stopping the agent after it read the wrong Notion docs). - **Example — Claude co-work:** progress/to-do list of done and upcoming steps; visibility into context and skills; all tool calls shown with inputs and outputs. - **Example — Manifold:** task progress showing completed and pending tasks; shows what it looked at and how it interpreted it. ## Pattern 3: Personalization - **Definition:** Give the agent the thoughts, systems, knowledge, principles, and patterns *you* would use for the task. - **Core claim:** optimize **speed to understanding**, not speed to outcome — many agents get this wrong; output not aligned with how the user wants it is "useless." Personalization enables the agent to "do the right thing and not just something." - **Example — Harvey:** "playbooks" capturing legal firms' methods/principles (e.g., contract review); agent then works the way the firm would. He qualifies: "I'm not a legal expert, but as I understand…" Harvey also uses **memory** built up across interactions. - **Example — Claude:** skills, connectors, connected systems to expand knowledge and personalization. ## Pattern 4: Reversibility - **Definition:** Ability to undo agent actions. - **Value:** **bounds the cost of mistakes** — knowing worst-case downside simplifies the ROI calculation, making users **bolder**, more risk-tolerant, and likelier to tackle higher-value tasks. - **Example — Cursor:** reversibility at multiple granularities: - Line-level accept/reject of changes - File-level acceptance - Rolling back conversation state (undo last messages and their changes) - Multiple parallel outputs with the same input using different models, knowingly keeping one - **Example — Harvey:** a Microsoft Word add-in using the native Word API for tracked changes, as a reviewer/editor would natively work in Word. ## Closing - Speaker notes it was "quite a lot for a short amount of time," hopes it was useful, invites follow-up questions. (Transcript ends with applause.) # Attribution Record: "What the Best Agents Share — Mardu Swanepoel, Flinn AI" ## Speaker Inventory | Speaker | Role | Evidence | |---|---|---| | **Mardu Swanepoel** | **Creator / primary presenter** | Named in title; delivers the entire talk in first person ("what I believe the best agents share", "when I personally develop agents"). Affiliation: **Flinn AI** (per title). | - No guests, narrators, or additional identified speakers appear in the transcript. - Audience presence is implied by **[applause]** at the end and one **[snorts]** mid-transcript (unattributed audience/other sound, not a speaking role). ## Referenced Third-Party Products / Claimed Sources The speaker cites several third-party agents as examples; these are the speaker's characterizations, not verified claims: - **Cursor** — focus modes (chat modes dropdown, planning mode, debug mode); reversibility (line-level accept/reject, file-level, conversation-state rollback, parallel multi-model outputs). - **Claude (co-work)** — transparent execution (progress/to-do list, visible tool calls with inputs/outputs). - **Manifold** — task progress visibility. - **Harvey** — playbooks for legal firms, memory, Word add-in using native Microsoft Word API for tracked changes. - **Claude** — skills and connectors for personalization. ## Quoted / Attributed Statements - **Pablo Picasso** — quote referenced: the speaker paraphrases Picasso's concept of "steal" (study deeply, make it your own, produce something better). The quote text itself is not reproduced verbatim in the transcript; the speaker's interpretation is recorded as his own. ## Promotional Segments - **None identified.** No promotional or sponsored segments are present in the transcript. The closing line ("Please reach out if there's more questions") is a contact invitation, not flagged promotion. ## Provenance Notes & Uncertainty - No curated metadata was supplied; speaker identity and role rest solely on the video title. - The speaker's stated goal — studying four agents ("steal" in Picasso's sense) — is the framing of the talk. - Examples shown refer to on-screen visuals ("on the right-hand side", "top left", etc.) not available in the transcript; descriptions of those visuals are the speaker's. - The speaker explicitly disclaims expertise on one point: **"I'm not a legal expert"** (regarding Harvey playbooks). - **Comment availability**: not indicated in supplied source material. \>\> All right. Um so, before I jump into sharing with you what I believe the best agents share, um I actually want to share a quote with you. And this is a quote that I actually keep quite close to me when I personally develop um agents. It's a quote that maybe a lot of you might be familiar with, although I do want to dive into a little bit of what Pablo Picasso meant when he said "steal" um in this quote. He didn't necessarily mean stealing in the sense of taking something physically that is not your own and presenting it as your own. Um but instead, he referred to going and looking at something, studying it deeply, really understanding it, and making it your own, and then using that to come up with something better and something unique that you wouldn't have been able to come up with have you not done this process. And that is really what I want to do today in this talk. I want to have a look at four of what I believe potentially to be some of the best agents that we have access to at the moment. Um go and study them deeply, understand what they do, and see what we can learn from them in order to ourselves actually build agents in a much better way. I'm going to have a look at four specific patterns that these agents use. Uh for each of these, I'm briefly going to touch on what exactly this pattern entails, importantly what is the value that it adds to you using them, and then thirdly, show you quickly how does that actually look in real life in these agents. The first one is what I call focus modes. And focus modes is really where we put the agent in a specific mode where we constrain the action and the input space. So, we go into a planning mode or a research mode. Um what do we get from this? Well, first of all, the biggest benefit is for us as engineers, we get the ability to improve the agent's output quality on the smaller constrained action space. Um we really can potentially go and say, "Let's drop a bunch of tools. Let's really refine our system prompt. Let's optimize our e-vals to do really well on this small space first before we just do anything." Secondly, what is also really valuable actually is from a user perspective. Um one thing in these kind of do anything, ask me anything agent um UIs is the fact that the user doesn't necessarily know what to do to get the best result out of the agent, and they also have very big expectations. So, by going into a specific mode, we actually say, "Let's align a little bit the user's expectations and also tailor their inputs and behavior um specific to this mode." Cursor does this really well. So, on the right-hand side, you can see the cursor chat interface, and you can very easily switch between different modes by simply selecting a drop-down. Um and each of these modes then has specific behaviors and expectations that it sets for the user. Um it then does very specific things. So, in the middle, we see planning mode. It actually doesn't write any code. It just comes up with a plan, and it asks you questions, and you should be fine with it because that's what you signed up for. In debug mode, it has a very specific like hypothesis-driven approach towards, "Okay, what are the potential issues with your with your code? Uh let's spin up a dedicated debug server and push logs there and actually figure it out." Um so, in my opinion, a really, really powerful way in which Cursor is using modes to actually do certain things really, really well. The second pattern is transparent execution. And what we're trying to do in this instance is really trying to make what the agent is doing and using and thinking extremely clear to the users. And the crux of what we're trying to achieve here is to shift from uh delegation to collaboration. To really making the user part of the process um and not just letting the agent come up with the end result. The benefits we're getting here is first of all, trust in the output. If I give you a task and you come back with just simply the results, I will have less of trust in the results than if you were to actually share with me your process, share the thoughts you had, what did you read, what did you assume, what were the things you actually aren't certain about. Um so, we really use this process of transparency to build trust in the eventual outcome that the agent comes up with. Um additionally, it also enables the user to intervene at an earlier point in time if it sees the agent is really doing the wrong thing, and thereby reducing waste. If at step two of the agent, we saw the agent has just read from, I don't know, Notion uh docs A and B, and I shouldn't have I wouldn't have done that, then we can very easily say, "Hey, I think let's stop and take a different approach." This is something that Claude co-work for me does quite well. Um top right, it has like a progress list or a to-do list of things that it has done and will be doing, so it makes it clear what's the step that it's about to take. It gives you a good idea of the context that it's using, the skills that it's drawing from. Um in terms of cool tool calls, it's actually showing you all of the tool calls that it's making and also the inputs and the outputs of those tool calls. Um and this really makes it quite clear to the user what is actually going on from an execution perspective of the agent. Manifold does something very similar. Uh you also have your task progress where you can see the tasks completed and to be done, and it also gives a very good idea of what it actually looked at and what it made of those things. The third pattern is personalization. Um and this is really where we try and give the agent the thoughts and systems and knowledge and principles and patterns that we would have used if we were to do the task ourselves. And fundamentally, what we're trying to get here is to optimize or rather increase the speed of understanding of the agent. And this is a point that I think quite a few agents doesn't really get right in the sense that they optimize for speed to outcome, but not speed to understanding in the sense that it's very easy to just generate an output for a user, but if it's not really in line with what the user wants in terms of how they wanted it, it's going to be useless. So, optimizing for speed to understanding in the sense of really understanding all of the nuances and implicit things from the user, how it would have approached it, is really critical for agent to do the right thing and not just something. Um personalization is for us a way of enabling a quicker speed to understanding for the agent and doing the right thing and not just something. This is something we get in various flavors in different agents. Um for me too, which is quite nice, is the one is Harvey. Harvey has this idea of a playbook, and a playbook is for legal firms um typically kind of the I'm not a legal expert, but as I understand, the methods and principles that they use to, for example, uh review a certain contract. And you can create these playbooks in Harvey, and the agent would then do it in the same way as what your legal firm would have done it. \>\> \[snorts\] \>\> Harvey also uses fairly common concept, which is memory, so it actually creates memories as we go along and as we instruct the agent, and it can then draw from that in subsequent interactions. Um Claude, like many others, also has the idea of skills and connectors and systems that you can connect to in order to um increase this knowledge base and and improve the personalization of your agent. And the last one is then reversibility. And reversibility is really the ability for the user to be able to reverse or undo the actions that the agent has done. Um and basically, the big thing that we are achieving from this is we're binding the cost of our mistakes. So, if we know what the worst-case outcome is or at least what the downside cost could be for me, it makes the ROI calculation much easier for me to actually say, "Happy if you go and do that" versus there could be fairly big consequences. This number one then results in users being bolder and much more prone to actually taking risks and tackling higher value um tasks and use cases for the agent to actually do. This is done really, really well for me by Cursor as well. Um Cursor actually enables this reversibility on different levels of granularity. So, top left, you can actually roll back or or choose on a line level what you want to accept or reject based on what the agent did. Bottom left, you can accept on a file level. Um bottom right, you can actually go back into certain points of your conversation state, so you can say, "We've now had this conversation, but actually the last three messages, all of the changes you've done, undo those and jump back." And then also, it actually gives the ability to really do multiple outputs with the same input in parallel using different models. And thereby, the user basically is knowingly saying, "We will undo all but ideally one of our outputs um in order to actually reach something that is valuable." So, Cursor makes it really, I would say, easy for you to not have many like much downside and experiment with things and try things out knowing that you can worst case just undo and and carry on. Um Harvey also does this quite well, and they actually use So, this product, it's a Microsoft Word add-in that runs in Microsoft Word, and they actually integrate with the native Word API in order to have this um change, you could say, doing of your changes and viewing of your changes in Microsoft Word as a reviewer or editor would natively using using Word. All right. Um thanks a lot. That was I think quite a lot for a short amount of time. I hope it was useful. Please reach out if there's more questions. \>\> \[applause\]