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

Full analysis

Your draft

Flo Crivello just launched Lindy Teammate, an AI "employee" that lives in Slack, and then said the quiet part loud: his whole product runs on DeepSeek, a Chinese open-weight model, and he thinks the US should ban it anyway. That's the tension worth chewing on. What does it mean for an engineering lead shipping agents into production when the cheapest path to viable unit economics runs straight through models you might not be allowed to use next year?

This is a Type 1 decision hiding inside a Type 2 shopping trip. Picking a primary model looks reversible. It isn't, once you've built an 85% cache hit rate and a prompt-optimization loop around one model family. What's actually being decided: how much of your cost structure you're willing to bolt to a supply source with real regulatory and geopolitical exposure. The forcing function is soft today but getting harder. Anthropic is already lobbying the position Crivello describes, and a $10K re-optimization bill per model swap tells you the exit isn't free.

The Skeptic. Crivello runs on DeepSeek and admits Lindy Teammate is in negative gross margin territory anyway. Sit with that. The 100× price advantage produces a less unprofitable product, not a profitable one. So the real claim here isn't "DeepSeek makes agents cheap," it's "agents at this token volume don't pencil out yet, even at near-zero model cost." For a PM: the model bill is the small line item. The architecture is burning tokens faster than any price cut can save you. And the ensemble trick he won't run because it breaks the cache? That's an admission the caching moat also fences you in.

The Compute Pragmatist. The economics are more fragile than the 100× headline. Cache hit rate at 85% versus 65% is roughly a 2× swing in cost, and Crivello says any system change breaks the cache. So the savings live on a knife edge you re-cut every deploy. The centenary tree, 100 children per node, two LLM calls to address roughly 2 billion tokens, is genuinely clever engineering. But it exists because the token bill is brutal. When your internal inference spend approaches payroll, you're not running a software company anymore, you're running a token refinery with an SLA.

The Open-Source Advocate. This is the story the closed labs don't want on the record. DeepSeek at roughly Sonnet 4.6 level for 100× less, Kimi K3 and GLM 5.2 in the same conversation as Opus. Open weights aren't 80% of the frontier here, they're at the frontier on the axes a product team cares about: capability per dollar, and you own the weights. Crivello's ban argument is, functionally, an admission that open Chinese models are winning on merit, so the counter has to be policy rather than product. For a builder: the leverage open weights give you is you can run them where you want and swap without a vendor's permission. The thing that erases that leverage isn't a better US model. It's a law.

The Researcher. Two claims deserve flagging. The 8× error-rate drop in week one came from an LLM-as-judge validator ensemble with a 10,000-token validator prompt, not from a smarter base model. That's a systems result, reproducible on any capable model, and it's the most transferable idea in the episode. The forward call, LoRA-per-user for continual learning within six months, is the shaky one. LoRA means cheaply fine-tuning a small adapter per user instead of retraining the whole model. Merging that with inference at production scale is unsolved, and "friends at labs are panicking" is not a benchmark.

The Enterprise Buyer. A shared, multiplayer agent that ingests all your Slack history, wikis, and meetings, built on a model whose weights answer "I can't talk about that" when you ask about Tiananmen. Procurement will ask exactly two questions. Where does our data go, and what happens to this contract if DeepSeek gets restricted. Crivello has handed every enterprise buyer the objection himself. The two-tier memory, private user layer plus public team layer, is the right architecture. But CCP censorship baked into the weights is a compliance line item no amount of clever memory design erases.

Where they part ways. The Open-Source Advocate sees open weights beating the frontier on cost-per-capability and wants that freedom protected. The Enterprise Buyer sees the same weights as an unsignable contract the moment a regulator moves. Both are right, which is the whole problem. Second split: the Compute Pragmatist treats the 100× DeepSeek discount as the thing that makes agents viable, while the Skeptic points out the product loses money regardless, so the model price was never the binding constraint. The architecture was.

What this actually hinges on for your team: is the DeepSeek cost advantage substantial enough to justify building around it, knowing a US restriction could vaporize it? The council leans toward no, don't bolt your cost structure to one model family whose availability is a policy coin-flip. The transferable wins here are model-agnostic. Validator ensembles cut error rates. Aggressive caching and tree-structured context cut spend. Two-tier memory makes shared agents work. None of that requires DeepSeek specifically.

De-risk it concretely. Measure your own cache hit rate and model it at 65%, because that's your cost if a forced migration breaks the cache. Build the $10K re-optimization loop as a standing capability you run on a schedule, so a mandated model swap is a Tuesday, not a rewrite. And run the validator-ensemble pattern now on whatever model you're already allowed to use. That's the part of Lindy's stack that pays off regardless of which way Washington jumps.

Prediction: By the end of Q1 2027, no US federal ban or licensing restriction on Chinese open-weight models (DeepSeek, Kimi, GLM) will be in force, and these models will remain freely downloadable and deployable by US companies.

Confidence: Medium. No bill exists, and open weights are already distributed and unrecallable.

Why: Crivello is arguing for a ban that does not exist, and even he retreats to "audits or insurance" as the realistic middle ground once Nathan Labenz pushes back, which tells you the outright ban has no near-term legislative path. The deeper mechanism is that open weights, once released, are already on thousands of machines and mirrors, so a ban would police deployment rather than access, a far heavier lift that Congress has shown no appetite to attempt on this timeline. The opposite outcome, a fast federal restriction, would require both new legislation and an enforcement mechanism for software that is already everywhere, and neither is close.

Revisit by 2027-03-31: We're right if US companies can still legally download and run DeepSeek, Kimi, or GLM with no federal licensing gate. We're wrong if any binding federal rule restricts their use before then.

That the frontier-adjacent capability now sits in open Chinese weights, and the only proposed counter is regulatory, is the part every US buyer should be pricing in. The engineering ideas travel. The supply assumption might not.