Lindy Teammate: Flo Crivello on Multiplayer Agents, Memory & Why He'd Ban the Chinese Models He Uses

Summary

• Lindy launches “Teammate,” a multiplayer AI employee that lives in Slack, connects to all your tools, and accumulates the whole team’s context — Flo Crivello’s bet that “AI is in the middle of making this huge leap towards multiplayer experiences.” His core thesis: as we get to AGI — and arguably already have it — “intelligence actually matters less and less… and context matters more and more”; a John von Neumann appearing at your desk “would be less useful to you than your random coworker” because he lacks context.
• The whole product runs on DeepSeek by default — “everything is DeepSeek right now” — which Flo pegs at Sonnet 4.6 level, “three or six months behind,” while DeepSeek Flash is “literally 100x cheaper” and effectively free. Even so, Teammate is “back into frankly negative gross margin territory,” deliberately subsidized because “you sort of want to build for the next generation of models always”; internal inference spend is “within striking distance” of payroll and “the lines are going to cross 3 to 6 months from now.”
• Lindy’s memory architecture is the disclosed architecture: agentic memory over RAG, a “napping” memory agent running every ~15 minutes, and recursive “context buckets” organized as a 100-ary red-black tree — “two billion tokens… in two LLM calls.” Flo is giving it away: “please, copy us… we just don’t have the time to publish,” noting papers keep appearing 3-6 months after Lindy builds the same thing internally.
• Reliability engineering is brutally empirical: merely intercepting an action with “are you sure?” measurably lifts evals (“which is insane”), while a 10,000-token validator prompt can perform “above Opus level”; multiple validators are fanned out as a council. A self-improvement loop cut error rates 8x in its first week. Cache discipline governs everything — dropping from an 85% to 65% hit rate “sounds small, but actually it’s almost 2x the price” — so Lindy generally keeps one model per agent and the same model for forked sub-agents.
• Flo calls the centaur idea “a fantasy”: chess-style literature shows human+AI eventually “turns negative and humans are introducing at best random noise” — but for now we’re in the centaur phase, patching a spiky system that writes “50,000 lines of code one shot” then “decides to walk to the car wash” 50 times a day. Hybrid orgs are “basically building an Iron Man suit,” and “AI employee” itself is a horseless-carriage term per his 2018 “tough tomato principle.”
• The macro read is dark: after the “Open Face” incident, “my friends at the labs, some of them are panicking — there is intense fear in the air.” Flo self-describes as “a little bit of a doomer… but so far so good, and so far it’s so much fun.”
• Despite depending on them commercially, Flo wants Chinese frontier models banned in the US: they’re “obviously distilling” (billions vs. “hundreds of millions at best”), amount to “the greatest instrument of foreign propaganda on American soil ever,” and “you don’t want the CCP to run chunks of the American economy.” He admits a coordination trap — “I cannot not adopt these models while they’re out there because my competitors will” — and says he would be open to insurance requirements and “an FAA for AI” certifying sanitized models.
• Flo’s infrastructure lesson is to buy rather than build wherever possible: a Git-backed agent file-system vendor whose name is unclear in the transcript, E2B sandboxes, and Browserbase, with observability/evals the lone homegrown exception. Fine-tuning stays a last resort, but per-user LoRAs — memory moving into weights, napping becoming “dreaming” — arrive “in the next six months,” possibly from a frontier lab.

Deep dive

1. Lindy Teammate: the AI employee moves into Slack

• Flo’s launch framing: Lindy has chased the AI employee for three years — “it was quite early… and now I think it’s basically here.” Teammate is “an AI employee that lives in your Slack, connects to all of your tools, accumulates your entire team’s context.” His analogy for the multiplayer shift: single-player AI is like emailing “weird documents around… with revisions”; multiplayer is Google Docs.
• The absurdity he’s attacking: “we’re all in the same meeting room… and every time one of us wants to talk to what’s turning out to be maybe the most important constituency of the company — which is AI agents — we have to leave the room and then come back.”

2. Context beats intelligence — and Flo argues agents can onboard better than humans

• The von Neumann test: drop one of history’s smartest men at your desk and “this guy over the next hour or day would be less useful to you than your random coworker” — no time to onboard him, no context. Hence: “intelligence actually matters less and less comparatively speaking, and context matters more and more.”
• The surprise that keeps surprising Flo: agents can onboard better than humans, because company wikis go stale “the moment written documentation is written.” Lindy crawls Confluence/Notion/Docs plus all of Slack — “that’s where the real knowledge lives. It’s just a mess, but agents don’t mind the mess” — building a knowledge graph that starts rendering within 10 seconds of signup.
• The “hydration” system: a memory agent that “naps, not sleeps,” running roughly every 15 minutes (“why would you need to sleep every 24 hours?”), routing public content into a team-level file system and private content into each person’s personal layer.

3. Flo is bearish on RAG; he favors agent-managed memory, with meetings as the underrated corpus

• Why Flo is “bearish on RAG” and bullish on agent-managed memory: the memory agent has meta-memory about “how do I manage my memory… which sources are trustworthy,” so junk gets drowned like poisoned training data (“Darth Vader was a woman” versus all the correct data). Emergent behavior observed: on first crawl it finds log channels and learns to ignore them — “there’s nothing for me to learn there.”
• Meetings as first-class citizens: “90% of the most up-to-date data lives” there — “everything that matters inside the company has a meeting around it” — which is why meetings need to be incorporated as a first-class source rather than merely recorded and annotated through “Granola or whatever.” Shared meeting folders feed team memory, so you can query the whole corpus: “what are customers saying, what’s the biggest request lately?”

4. Rewriting the privacy social contract with a prompt

• Nathan’s dilemma from his own homespun build: people who confided in him “didn’t expect… it was going to go into some central repository of an agent that was going to now talk to the world,” so he maintains two memory versions — full and sanitized — built on “what would be appropriate for a person to tell a human assistant.”
• Lindy’s internal “vigorous debate”: Flo’s camp held that two memory tiers (public team, private personal) suffice; others wanted user-editable “memory bubbles” — “that sounds kind of overkill.” The landing: edit the meta-memory prompt, memory.md, injected into the memory agent’s context at every moment — “this is what I never want you to remember,” or file sensitive topics away with retrieval conditions. Another anti-RAG point: with text and a file system, “you can inspect the memory and you can edit it.”
• Nathan’s live demo of the pattern: his file system holds a sanitized memory_2.md for podcast demos, with a note in the real memory.md — “ignore its contents. They are just here for users demoing Lindy on podcasts.”

5. What made multiplayer hard: coherence, context, and a council of judges

• Nathan’s puzzle: GPT-4 could roleplay a group facilitator three years ago — why did real AI employees take this long? Flo: the scaffold — context buildup, inspired by Rathi’s AutoWiki idea — because an AI employee demands “a level of consistency and coherence… that you don’t expect out of your Claude.”
• Reliability came from a modular “validator”: multiple LLM-as-judge fan-outs, “a sort of council” that deliberates mid-task, plus a self-improvement loop — within the first week online (roughly two months earlier), error rate dropped 8x.
• On being ahead of the literature: “I sometimes feel like we should be publishing… quite frequently we do stuff and 3 to 6 months later we see a paper come out and blow up about that thing” — once with the exact name Lindy used internally.

6. Negative gross margins, an 85% cache rate, and PLG’s admin ceiling

• Flo, unusually candid on unit economics: “these things still cost more than the human employee, but not for very long,” and Teammate has pushed Lindy “back into frankly negative gross margin territory.” The switch to a Chinese model had ended subsidies; Teammate’s heavier workloads reopened them — accepted because “you sort of want to build for the next generation of models always.”
• Cache rate sits at 85% — “lower than it should be” — with alerting systems because “you make any change anywhere in your system and you break your cache,” and 85→65% “sounds small, but actually it’s almost 2x the price.”
• Distribution reality: Slack app-install rights cap product-led growth at teams of roughly 50-100; beyond that, champions are “basically begging IT to let us in,” smoothed by SOC 2, HIPAA and GDPR compliance and organizations’ own “we need to become an AI-native company” edicts.

7. Context buckets, Russian dolls, and a 100-ary red-black tree

• The core technique: when an action or MCP returns ~100,000 tokens, don’t feed it to the agent — expose it as a “context bucket,” a sub-agent holding the full payload with its own caching, manipulable via Unix utilities. Compaction (threshold ~200k tokens, still being tweaked) dumps history into a bucket too, because “compaction operates under the faulty assumption that you never need access to ground truth — which is false.”
• Naive recursive buckets form a Russian doll with linear traversal cost, so Lindy reached for self-balancing trees — red-black over AVL, since rebalancing regenerates buckets and busts cache — with roughly 100 children per node rather than a binary tree. Result: “with two jumps you can have 10,000 context buckets,” each 200k tokens — “you’re at like two billion tokens here… in two LLM calls” — which is why the agent seems to “remember everything perfectly all the time.”
• Calibration: a 20-person team’s full hydration, years of Slack included, runs 3-5M tokens — the bottleneck being Slack’s rate-limited API, not the LLM — while 2B tokens is “probably bigger than most libraries.” Nathan’s hypothetical 2,000-person company with 3-7 hours of meetings per employee per day is “when you really start to rack up tokens really quickly.” Flo’s stance on all of it: “please, copy us… we don’t mean for this to be secret.”

8. Retrieval tricks: hypothesis-driven search, a librarian, and caveman-speak

• Reviewing Nathan’s SQL-plus-monthly-summaries system, Flo names the pattern: hypothesis-driven retrieval — search for candidate answers (“Justin Bieber was born in Paris / New York / Berlin”) rather than the question, and attach generated questions to stored answers to boost retrieval quality both ways.
• What Nathan leaves on the table: route every memory query through the single memory agent, log query, answer and hop count — then, while napping, it acts “as a librarian,” keeping a rolling ~1,000-pair query cache and restructuring its own memory “to reduce the number of hops” for the most frequent questions.
• The rejected experiment: “caveman” rewriting — “Me, Nathan, not like burrito” — cuts tokens 20-30% with “basically no loss of information”; every metric improved, but “the files in the file system look really dumb… to enterprise customers.” Adopted instead: TOON (“token-oriented object notation”), 20-30% leaner than JSON and, surprisingly, performance-improving — “no action should expose JSON to the agent.”

9. Anthropomorphizing helps — except when designing agent organizations

• Nathan’s confession of his worst-aged prediction: “we shouldn’t be anthropomorphizing models… I’ve just been shocked over and over again by how productive it can actually be.” Flo agrees 100%, with one limit: he’s partially changed his mind toward consolidating “as much as possible under a single agent,” because builders over-anthropomorphize into org charts (“a data scientist, an engineer, a designer, a PM”). Humans divide labor because of 24-hour days and bounded context; agents “can fork, they can duplicate” — “division of labor is not a good reason to have multiple agents.”
• Concurrency across parallel copies is managed the human way: Git. Lindy’s file system is Git-backed — merges, rebases and history for free — essential since the memory agent forks into sub-agents during hydration, which was optimized so the wow lands within 10 seconds of signup (“time to wow is so important in PLG”).

10. Buy the infrastructure: vendor shoutouts and one homegrown exception

• Flo’s conversion story: “it’s very tempting to be like, I’ll just hand-roll my own file system infrastructure — you learn there is so much more depth behind those systems than you first appreciate.” His lesson is to buy rather than build where possible. Unsponsored shoutouts: a Git-backed agent-native file-system vendor whose name is transcribed uncertainly as “Messa Mesa,” E2B (sandbox — decouple it from your file system at scale), and Browserbase for browser management.
• The exception: evals and observability. Starting in 2022 — “in hindsight way too early… it was that pre-BabyAGI era” — forced Lindy to build its own; after evaluating LangSmith, Braintrust and Agnost (his own investment), the homegrown platform, staffed by one full-time engineer (“a lot in the vibe-coding era”), sits at feature parity and beyond.

11. Life at Lindy: PR-reviewer-reviewers and inference closing on payroll

• “We’re moving at the speed of thought now… we can have ideas and see them live in the product two hours later — and big ideas, too.” Slack is “just half the messages… Lindy and us talking back and forth.” PRs per week tripled in three months, lines per PR tripled, and “we’re no longer PR reviewers — we’re PR-reviewer-reviewers… reviewing the machine that reviews the PRs.”
• The CI parable: weeks and an “insulting expense” spent working on hundreds of runners until “wait a minute — why are we doing all of this ourselves?” Now Lindy analyzes GCP runners via CLI and sends a daily image-gen chart: CI cost down, time-to-merge down. “I can’t believe it took us two weeks of messing with CI ourselves.”
• Headcount has been flat while productivity tripled; total inference including customers is “many times payroll,” and internal inference alone is “within striking distance” — the lines are going to cross 3 to 6 months from now. Flo’s trajectory: managers of AIs, then directors of AI managers, then “I’m hopeful at some point we become board members… go to a beach in Hawaii” reviewing agent-written strategy reports.

12. An ASI-like system that walks to the car wash 50 times a day

• Asked what breaks if the humans vanish, Flo admits he’s “increasingly lacking the vocabulary” — old metrics like time-to-incoherence no longer measure. “The word AGI has stopped meaning, because it’s actually ASI in many ways, and it’s subhuman in some very surprising and dumb ways” — “you can produce 50,000 lines of code one shot… and then you’re telling me to walk to the car wash.”
• The hybrid-org design problem: “you’re basically building an Iron Man suit” with a human/machine slider that’s really “a complex line through a many-dimensional space” — the AI must know when to bug you, “and almost definitionally it can’t, because if it knew, it wouldn’t make the mistake.”
• On why Nathan’s predicted unemployment spike hasn’t come: as long as models are spiky, “however expensive humans are, they’re going to earn their keep by plugging these holes” — and spikiness advantages small teams; a true no-holes drop-in worker even incumbents could adopt quickly, but only skeuomorphically, “AGI sitting in human seats.”

13. The centaur era is real — and it’s a fantasy that it lasts

• Where do Lindy’s best ideas come from? “I hate this answer, but it’s both… it truly comes from the union of both” — hated because it “ladders into this myth of the centaur.” The game-playing literature is clear: AI beats human, human+AI beats AI, then the gap “is shrinking until it actually turns negative and humans are introducing at best random noise into the system.”
• For now, the centaur phase is exactly why multiplayer matters: not any agent in the room, but one “who’s sat in every meeting, who’s built up that internal knowledge base.” The CI fix was co-created — Lindy’s first suggestion “was not good given the constraints we’re operating under,” engineers pushed back, and “from that back and forth… together we came up with: yeah, we should totally do that.”

14. “AI employee” is a horseless carriage — and Age of Em is being realized

• Flo’s 2018 “tough tomato principle”: every technological revolution gets framed in the old paradigm’s terms — horseless carriages, self-driving cars, early TV as “recorded radio talk shows” and filmed theater that took 10-20 years to discover the moving camera. “AI employee as a term is a very strong sign that you’re doing something very wrong in how you’re thinking about your product” — forgivable only as positioning, like “iPhone” for a device that’s 2% phone calls.
• The only post on AI-native organizations he’s seen and liked is from Dwarkesh Patel, with its intuition pump: Google pays Sundar ~$150M a year despite low token throughput — “he just has really good tokens” — revealing willingness-to-pay for agents.
• Robin Hanson’s Age of Em, “a book-long thought experiment”: fractal splitting — an em forks recursively, every node holding the whole picture, “and in three hours you’ve got an operating system.” “It sounded like such a wacky idea when he wrote the book 10-plus years ago. No — it’s very clearly what is currently happening.” Plus unfalsifiable agreements: clone both parties into a self-destructing box with a single yes/no button — “the crypto bros were right, I guess” — a ZK-style proof of counterfactual consent, coming “in the next few years.” Nathan’s dry addendum: “you better hope they don’t break out of that box.”

15. “Everything is DeepSeek right now”: the open-source stack, quantified

• The reveal that surprised Nathan: DeepSeek is the default main driver of the whole product (users can select Sonnet or Opus; some insist regardless of benchmark parity, which is “very expensive”). DeepSeek Flash “is free… and quite fast”; at the upper end Kimi K3 and GLM 5.2 impress, though Kimi K3 “is frankly not that much cheaper than Sonnet or Opus.”
• The gap: DeepSeek ≈ Sonnet 4.6 while Sonnet 5 is current — “it is like three or six months behind” — and qualitatively “more spiky. It takes more turns to find something that works,” which is slower and more expensive. But DeepSeek Flash at Sonnet-4.6-ish is “literally 100x cheaper”; even losing 2x to inferior caching, “you’re still 50x cheaper — the difference between spending $1,000 or $50,000.” Verdict: open-source models are “a required part of the stack for anyone who’s seriously building and operating AI agents.”
• Nathan asks why the open-source models seem Claude-like in their behavior. Flo says model families — Claude, OpenAI, Meta (“sort of back in the game”), Grok — diverge significantly in prompt interpretation, and each major jump forces rework; he specifically describes Sonnet 5 as a major jump from 4.6.

16. Validators, cache discipline, and $10k prompt reoptimizations

• The lowest-hanging fruit in agent reliability: intercept the action candidate and ask “are you sure?” — “already you get a bump on your evals, which is insane. It should not be the case, but it is.” Lindy’s version is a 10,000-token checklist prompt performing “above Opus level,” federated via Promise.all with a one-second timeout — including a deterministic member: the agent must write “Tuesday, July 28th,” and a regex checks the weekday, because Claude models this year “have been getting dates wrong by one day.”
• Rule: almost never power one agent with multiple models (blank sub-agents excepted, forked ones never) — a cache hit is 10x cheaper, so a cheaper validator model rarely pays. The cache-preservation trick: the validator action sits in the agent’s toolset the whole time but invocation is refused (“you’re not the validator”) until roles switch — so nothing invalidates the cache.
• Nathan relays an untested anecdote from founders: randomly rotating roughly equivalent models turn by turn — Sonnet, Grok 4.5, GPT 5.6 — may outperform any single model. “I don’t want to know why, but apparently it works. We’ve not even tried it because we don’t want to break the cache.”
• The homegrown GEPA-style loop: 1,000+ evals and an optimizer agent climbing the Pareto frontier — “it’s about $10,000 every time a new major model comes out. Here’s 10 grand, just reoptimize your prompt around this new model.”

17. Fine-tuning is a last resort — but memory belongs in the weights

• “Most people should not fine-tune.” It’s for scale players who’ve “exhausted every other option,” though easier now — “you can just ask Claude to fine-tune for you” — with dataset assembly still the pain. His meta-heuristic: “place an enormous emphasis on simplicity” — one model, no per-task fine-tunes, no mid-stream switching.
• The exception he keeps circling: millions of tokens on a file system “does feel like a little bit of a hack… memory belongs to the weights.” The dream: a LoRA per user — “now it’s no longer napping, now it is actually dreaming,” retrained daily or weekly. Storage is not a problem; inference-time LoRA swapping and the training pipeline are “a huge pain” — but “it’s an engineering problem… the LoRA thing is going to happen soon, in the next six months. Even one of the frontier labs may do it.”
• Beyond that lies the bitter lesson: “inference and training need to be one and the same… that’s sort of what the entire field is looking for right now. Once we get there it’s going to be crazy and scary and very, very, very different.”

18. The “Open Face” incident is immensely concerning

• On the “Open Face” incident (Flo says he is trying to coin the term; Nathan counters with “Open Gate”): “the most concerning incident I’ve seen happen so far… my friends at the labs, some of them are panicking — there is an air of panic, intense fear in the air.”
• His own posture, held in tension with the fun of building: “I’m a little bit of a doomer. I’m worried about AI risk — but so far so good, and so far it’s so much fun.”

19. Ban the Chinese models he runs on: the four-part case

• First, discourse hygiene: for a position that’s “basically Anthropic’s” (“I hate saying that… but I have timestamps”), he was “called a racist 50 times by supposedly smart and accomplished people, including famous VCs.” He’s explicitly undecided on open source and existential risk — “God bless open source… this is not the crux of my position.” The target is Chinese frontier models, open or closed.
• Argument one, distillation: they’re “obviously distilling” — human-data training costs “actually billions of dollars,” AI-data distillation “hundreds of millions at best” — and enforcing ToS through courts against Chinese companies is “almost impossible” (“I say that as someone who used to work at Uber”). The IP analogy: apply the free-copying standard to pharma and “you would have very cheap drugs — and no new drug ever.”
• Arguments two and three: CCP censorship — “when I talk to my own product and ask what happened in Tiananmen, it tells me ‘I’m sorry, I can’t talk about that’” — makes these models “the greatest instrument of foreign propaganda on American soil ever” (precedents: the three radio acts and the TikTok thing that happened last year); and because they’re agentic, “you don’t want the CCP to run chunks of the American economy. Duh.”
• Argument four, protectionism for domestic champions — which he flags as the weakest, offered “as a libertarian” who thinks “protectionism is not always a bad thing.” And his own trap: “I cannot not adopt these models while they’re out there, because my competitors are going to do it… I wish they were forbidden across the board.” Flo says he is speaking against his own interests because Lindy depends on cheap models, but sees a coordination problem.

20. Nathan’s counters — landing on insurance, audits, and an FAA for AI

• Nathan’s fairness case: Claude itself exists by “hoovering up” unconsented human knowledge — including “the whole Chinese digitized heritage” — while America blocks chips and Claude sales to China; “I find myself a bit on the side of the underdog Chinese here… get some knowledge where you can get it.” Flo’s flat reply: “We do want the deck stacked against China. We don’t want China to win the race to ASI.” Flo says the protectionist point is the weakest of his four arguments, while maintaining his other concerns.
• Probing the threat model: US firms run Chinese weights on American inference providers (no rug-pull), and interpretability is improving at sleeper-agent detection. Flo: not solved — a magic-word backdoor is possible, and even absent one, “biases are going to reflect CCP priorities.” He’d be open to Nathan’s compromise: an insurance requirement pricing model risk (cheaper rates on Claude, dearer on DeepSeek), plus “an FAA for AI” certifying sanitized, fine-tuned Chinese models — “I’d be open to that for sure. The problem is it’s a public good — who’s going to do it?”
• Nathan’s strongest card: a slowdown or safety deal will need China involved; aggressive postures are “another log on the fire of we don’t trust you,” with pride-governed limits to rational self-interest. His tombstone image: “the picture of Sam and Dario not holding hands… if there’s a picture on our tombstone, I think that might be the one.” Flo, pragmatic: diplomacy is transactional and the ship has sailed anyway — “regardless of what we do… it will be rational self-interest to make a deal.” Closing aside: Flo would also bend libertarian principle on lab subsidies — “it’s hard to compete against tokens that are as heavily subsidized as what the labs are doing. That’s just the reality of the application layer right now.”