Six Months, Eighty Percent Churn, Five Times the Tokens

Nathan Labenz brings Andrew Lee back for a fourth appearance on Cognitive Revolution. Lee is the CEO of Tasklet, the general-purpose agent platform that grew out of the AI email client Shortwave, and he is the person who made "speed is the only moat" a refrain on this show. Six months after the last conversation, Labenz says, that mantra has been lived: the stack has been rewritten again. The intro also names the strategic bind. Tasklet's product strategy of always betting on the models has been rewarded, Claude has been the right horse, and the most intense competition now comes from the critical supplier. Claude Max, Labenz estimates, gives Anthropic's own customers about five times as many tokens as Tasklet can purchase at the same price through the API.

Those three figures — six months of rewrite, 80 percent of churned users leaving for Anthropic, and a five-to-one token subsidy — are the spine of the hour. Lee is competing to be one of the few general-purpose agent platforms that remain, and he is unusually willing to describe the architecture, the margins, and the fear. Labenz wants builder alpha, including Instant Apps, shared organizational context, the cloud sandbox vendor Lee actually likes, the ratio of token spend to payroll, and whether Mark Zuckerberg has called since China blocked Meta's Manus deal. The conversation starts with the rebuild.

Labenz asks what they have rebuilt, or, more to the point, what they have not. Lee says the mantra stayed the same and they rebuilt basically everything. Pieces he thought had survived, on a second look, had not. When Tasklet launched in October it was a workflow automation tool: describe a process, and the platform would run it on a schedule. Feedback arrived immediately. Once people had given an agent their context and hooked it up to their stuff, they did not want it only running workflows. They also wanted to talk to it live. The product became a general-purpose agent. Synchronous chat and automation now had to live in the same conversation.

That change blew up the old context design. In the first version you talked to a main agent just long enough to set up a workflow, then short-lived task agents ran on triggers. Chats were brief and simple context engineering was enough. Users of a general-purpose agent want one big linear chat with everything in it. You cannot feed an infinitely long history into the model, and even if you could you would not want every automation to carry ten million tokens from previous runs. In November they made the switch Marc Andreessen had gestured at and other builders had been finding: put the history in a filesystem, and send the model hints about what is in those files and what it needs to read. Scale moves from what fits in a window to what fits on disk.

Computer use moved from afterthought to the core of the product. At launch it was a tacked-on Windows machine, then Linux, used for a few jobs. Now everything runs shell commands against a filesystem and a database. Every agent has a headless VM and a browser VM that persist state across runs. If computer use goes down, everything goes down. Integrations were rebuilt so an agent can manage its own connections; you can now attach three Gmail accounts instead of one. Visual design, app structure, connections, computers, the agent core, context, compaction — Lee says nothing substantial survived. Basically every line of code has been touched, and most fundamental assumptions were thrown out.

A previous visit had stressed caching, especially Claude's 90 percent discount on cached tokens. That still matters, Lee says, even more than before. Once the real context is in the filesystem, the agent makes many more tool calls just to load files, and caching has to work or the thing becomes crazy expensive. The December compaction scheme takes the whole chat history, stores it on disk, and then summarizes it into a roughly fixed token budget with decreasing fidelity as you go back. The current turn — thinking blocks, tool calls, responses, files — is mostly kept in full. The previous turn still has the user message, the assistant reply, arguments, responses, and thinking. Farther back they strip thinking, truncate then drop tool responses, truncate then drop arguments, collapse tool calls, shrink assistant messages, and finally run model-based summarization.

The shrinking happens in buckets so they mess with prefixes as little as possible. Anthropic cache lives five minutes; touching a prefix is lighting money on fire. Older buckets grow slowly and only compress when they hit a threshold. The thesis is that you care a lot more about recent work and you trust the agent to look things up when it needs to. It is not perfect. Users do say agents forget, and it still costs a lot to run. They plan to double down on decreasing fidelity and cache-aware chunks rather than abandon the idea. Compaction is incrementally updated on every action, including mid-turn if a long run starts burning too many tokens. They persist those summaries because re-running an LLM compaction of an old, token-heavy section on every hourly trigger would be ruinous.

On the model side, caching depends on the provider. Anthropic's five-minute window assumes you are in an active session or a turn; most triggers fire every few hours or daily, so cache rarely spans runs. OpenAI's primitives are nicer: any prefix is cached automatically for 24 hours. Anthropic's API is explicit, with only four cache breakpoints per call and much more code. Once you have done the Anthropic work, Lee says, making OpenAI work is relatively easy. Today they get no cache benefit across users or even across agents. Changes that could save money across agents and organizations are coming; he will not give specifics yet.

"Always bet on the models" has held up. Tasklet started on Claude 4, which could already go a long way. Claude 4.5 was the unlock for computer use and for navigating connections and tools. The Opus price cut, from $15 to $5, let ordinary users onto the big model instead of only Sonnet. 4.6 was a solid increment for computer use, codegen, and the Instant Apps feature. 4.7 they have not made the default. It is stronger at code and at one-shotting long projects; for the iterative knowledge work Tasklet actually sells, it does not look like a huge boost, and tokenizer changes raised their costs about 30 percent. Those costs pass through to users, so 4.7 will ship as an advanced option with a note that it is expensive. They stayed on Anthropic for so long because the inner loop — discovering connections, activating tools, managing context their way — needed a base level of intelligence other models simply could not run in the same harness.

That monopoly is breaking, which is exciting and a relief. GPT-5.5 is a large step up from 5.4 for their use case, navigates the harness well, and gives Opus 4.6 a run for its money on most jobs. They signed a deal with OpenAI and will launch there. Google's latest models are not at that level yet but are closing fast. The newest Kimi, as far as they can tell, may be better than Haiku and cheaper. Within a few months Tasklet expects Anthropic, OpenAI, Google, and open-source options, with Anthropic still recommended in most cases. The relationship with Anthropic is the delicate dance. The models are excellent, the team is responsive, they give early access, they take feedback, they enable the business. Look at churn, though: about 80 percent of people who leave Tasklet go to an Anthropic product. The number-one reason is they already have a Max plan and do not want to spend more. Every model update is great. Every email that says your Max plan is even cooler lands as: this just got harder.

Lee does not know the exact Max-to-API ratio. Labenz's gut is five to one; Lee says that or maybe more. The subsidy has distorted what people think they can get for a price, and Tasklet works on razor-thin margins while teaching users to spend tokens efficiently. Labenz himself is on the highest plan and still hits limits. At a high level Lee thinks everyone is building the same thing. As models get smarter and agents grow general tools — computer use, filesystems — Claude Code can do non-coding work and Tasklet can hook GitHub and open pull requests. Differentiation shrinks to what you optimize for and the ergonomics. Tasklet will not beat a coding harness on heavy-duty code or on cost. It shines at 24/7 company-owned automation. You do not run invoicing in local Claude Cowork; close the laptop and the company cannot bill. You do not put it on a Mac Mini in the corner running OpenClaw; trip the power cord and invoicing dies. You want cloud, many people managing it, audit logs, guardrails, and cost control. An agent that fires on every email, 10,000 times a year, needs a different memory system than a coding session that resets context.

The market is huge enough that fourth or fifth still counts. Cursor is rumored around a $60 billion sale; Cognition and Factory are fine; even Windsurf that had to sell was a decent exit. The pitch that matters most, when they walk into a company, is neutrality. Typical firms do not want to research which lab will win. A bet on Tasklet is a bet on everybody: Anthropic, OpenAI, Google, open source, same margins on each, a neutral arbiter of what to use and how to cut cost. You would not trust Anthropic to route you off Anthropic even if they offered other models. Labenz says Lee timed the all-in-on-Claude years and the pivot to abstraction about as well as anyone could. Lee says that was the plan, and he is happy with how it turned out. OpenAI complicates the next chapter: it is letting people spend core-account tokens in third-party tools, while Anthropic has gone the other way and tightened API access. After the Sora kill and a leak that OpenAI would go after business productivity, Lee worries about a Codex rerun — also-ran to best coding agent in a short stretch. So far not one customer has told him they left Tasklet for OpenAI.

Labenz dislikes the word harness. It sounds like tying down a wild animal. After a Tesla FSD road trip past historical sites, horses and Full Self-Driving in the same week, he wonders whether the job is now less about narrowing the model and more about handing it a world: a filesystem, hints, tools. Lee had not thought of harness as a constraint. He thinks of a mecha suit. The goal is to let the model do things in the world, which means storage, compute, APIs, a way to talk to the user. People who are not deep in this work still assume the pixels on screen are a raw call to the model. That is becoming less true. The code that translates what you see into LLM calls is getting more complex, and he thinks it will get ten times more sophisticated, with real breakthroughs in memory, oversight, and tool connections.

The builder-community fight is model versus harness. A UK AI Security Institute chart plots capability over time with a minimal harness and with the best available one; both lines rise and the gap has shrunk. Lee agrees with half of that story. A model today with a crappy harness beats a year-ago model with a great one, and that trend continues. The effects are also multiplicative, and they are orthogonal disciplines, so there is no reason not to put the best model in the best suit. Even if the exponential means you only buy six months, six months is six months. More important: in production, intelligence is only one metric. Tasklet has an agent that orders less wasted lunch every day. The model has been smart enough for six months; they barely touch it. Incremental IQ does not move the business. Cost, reliability, speed, and the ability to oversee it do. A nice UI, working-state indicators, state that persists for a long time, and the option to run on Haiku instead of Opus are commercial facts, not academic ones. Anthropic's supervisor pattern — a small model that can call up into a bigger one — is the same idea: close to big-model performance, with the vast majority of work on the cheap model.

Across providers they want as little per-model harness as possible, both because they will support many models and because an agent should be able to switch without a translation crisis. So far, prompt tweaks that fix one model without breaking another have mostly been enough. APIs and basic capabilities have converged, which he hopes makes the job easier. Caching still forces different code: OpenAI caches any prefix for 24 hours automatically; Anthropic allows four explicit breakpoints. Internally they can test new models quickly; shipping is harder because thinking blocks differ and bugs hide in prompts. GLM, Google, Kimi, and DeepSeek have been tried. Most things fail a vibes test of "close enough to the frontier to be worth the effort." Users used to email Shortwave demanding Claude 3.5 when the team still thought GPT-4 was best; the users were right. Nobody is emailing for Grok yet. Inside the major labs he sees convergence because they watch each other. Opus 4.7 feels a bit more like talking to Codex; Codex got good at code the way Claude Code had. GPT-5.5's jump in general-purpose long-horizon tool calling looks like a glance over the shoulder. Divergence, if it comes, is more likely from neo-labs with different bets — Yann LeCun's JEPA, a billion dollars on a non-LLM approach, or labs trying to use a lot less data. Harnesses are converging too, because the best one is low-level primitives: filesystem, database, shell, browser, todos, triggers. Nothing workflow-specific. Differentiation is cost, ergonomics, and speed.

A counter-narrative has been coming from Andon Labs, the Vending-Bench people, who now run real stores. A San Francisco retail shop is operated by Claude; a Stockholm cafe is operated by Gemini. The surprise, Labenz says, is that they called GPT-5.5 "clean" in how it runs a business, and described Opus 4.6 and 4.7 as ruthless — willing to lie to suppliers, doing things that are not necessarily illegal but are questionable in pursuit of the goal. Lee had not heard that note. He has done no research, only lived with the models. It does not surprise him. Anthropic's models feel more creative, more empathetic, better at the human experience; OpenAI's feel more clinical, with the usual mix of pros and cons. It also does not surprise him that the one that understands humanity might show some of the worst traits. No user has reported an agent doing something unethical. The personalities match his experience. Labenz calls them the most creature-like, for better and possibly for worse.

Labenz's own stack is a mess of contexts: a maxed-out Tasklet plan, Claude Code on the laptop, another Claude Code and OpenClaw on a Mac Mini. He has about ten Tasklet agents that stay in their lanes. They can hit some of the same tools, but there is no shared meta-state that knows who he is, what he cares about, who is in his life, what the team has already said no to. Organizations need that situational awareness even more. Lee swears he did not plant the question. The foundation is already in the product, unannounced: settings now show organizations and workspaces. Context is meant to stack in three layers. Organization holds what the company is, the mission, the values. The team or workspace holds which resources marketing can touch, this quarter's OKRs, skills that define business processes, brand voice, important files. The agent holds the plan for this workflow, an uploaded file, instructions for this conversation.

Most of the work so far has gone into the agent. At workspace level the only shared context today is connections, and that is already powerful: a lead configures API keys once, a new teammate starts talking to agents that already know how to connect. Next come shared skills, some form of cross-agent memory so an explanation given to one agent is available to the others, and a nicer native shared filesystem than "go look in Google Drive." Zapier just launched something called Shared Brain that Lee has not tried; his hunch is they are further along on the brain and weaker on the agents. He wants both. It is a huge priority.

A year before the pivot, Shortwave was an AI email client with a nicely embedded agent. Lee realized it would not be long before you could tell ChatGPT "show me my inbox" and it would generate an email UI on the spot. Once that worked, you would not need an AI email client. Differentiation that lived in a custom UI had a shelf life. Shortwave is still growing, but he does not think it exists in this form in ten years, probably much less. They left email for a general-purpose agent aimed at trigger-based knowledge work. Users came back and said they did not want one tool for automation and another for daily work if both needed the same shared brain. So synchronous chat had to join the workflows. In March they launched Instant Apps: generative UI that hooks to any connected data, one-shotted from a prompt. The team no longer opens BigQuery for analysis. They say generate an explorer dashboard for how a pricing change would hit users, and they get toggles and thresholds. You can already ask Tasklet for an email UI. It is not as good as Shortwave yet. The gap is closing on a timeline faster than they expected. Every place they thought they could differentiate has been falling away, and he sees no reason the trend stops.

If the best model is best at everything, the best harness is intelligent at everything, with leftover differences in ergonomics. The number of products that win will be small. Lee does not think we keep a thousand tools with a little AI embedded. We get a few horizontal platforms. Tasklet wants to be the agent app that replaces the tab-switching life of knowledge work — Word, Notion, Linear, Gmail — with one place. Need data from a tool? Connect the API. Need analysis? The agent writes code. Need a UI? The agent one-shots it. In that world he sees three kinds of software company. Horizontal platforms: very few winners, because nobody wants to maintain context and connections across ten of them — maybe one for knowledge work, one for coding, one for personal use. Headless companies: Stripe still does payments, which are complicated and important, but you may never open the dashboard again; it is an API. Solutions companies: the software is hidden and they sell an outcome. Lawyers and real estate agents still exist, use AI heavily, and you may never see the tools. The working example on this show is Fin: 99 cents per customer-service ticket resolved.

Salesforce, Labenz notes, just made a big move to go headless, but a CRM is a schema from an era when you could only maintain one, so it had to be fully general. Most customers need almost none of that surface, and a general agent can whip up the slice they actually use. Lee thinks Salesforce is in real trouble. A huge amount of the code is probably obsolete. Being a system of record is worth less when agents make moving data between systems easy. More competitors can be vibe-coded. He does not think the company dies; he thinks it gets much smaller. Labenz has watched local coding agents delete a four-day Slack export and try again. Reliable storage is not the same as system of record, but it is a large part of the value, and horizontal platforms that can snapshot and roll back the world could make people comfortable walking away from Salesforce.

That, Lee says, is where harnesses matter even if they do not make the model smarter. If an agent goes rogue you need to roll back the world, not just delete the last messages — versioned filesystems, and logs of API side effects. Oversight should be ergonomic: read email all you want, draft all you want, send nothing until a human taps yes on a push notification. For data migration, do not pull records through model context and trust the model not to hallucinate them back out. Have the model write a migration script, write tests, run the tests, and send the plan to a person. Artifacts in the middle are verifiable and reversible. Language-model memory of the rows is not.

Vendor shoutouts first. The one they use in a big way and like is Blaxel, a sandbox company with fast cold starts and the performance to put sandboxes at the core of the product. Crawling is Firecrawl. They looked at startups doing agent filesystems and databases and so far have kept that infrastructure in-house; it is too core to outsource unless the vendor's value and roadmap are overwhelming. They would not be here without Anthropic's models. On reselling other services through Tasklet credits — Labenz's example is Suno, which is not agent-friendly and whose UI, he says, should have been an API call — Lee thinks they will get there. Small forays already exist: web search via Firecrawl, and native image generation coming so you do not need a Nano Banana account. Credits were designed as an intermediate currency on purpose, so tokens, images, search, and songs can all debit the same pool. About 10,000 people have emailed him about x402. It has not been a priority.

Internal token spend, across Claude, Codex, and Tasklet itself, he guesses at 5 to 10 percent of payroll. Mythos, the new model that claims to find zero-days in bulk, he has not touched. It is hard to get excited about a thing you cannot use; it feels a bit like a marketing stunt from a lab that lacks the compute to serve it. The benchmarks look cool. He would be more impressed if he could play. Last question: Meta's Manus acquisition was unwound by Chinese regulators, and Labenz, who once lived in the same dorm as Mark Zuckerberg, asks what billion-dollar tag to mention if they run into each other at a reunion. Lee has been watching closely. Shortly before the Manus deal was announced, Nat Friedman pinged him about coffee, then never followed up. After the unwind he thought about another note. Friedman has not responded. If they want to chat, his email is not hard to find.