I stopped running agent teams and moved everything to Solo | Building Products [Mischa Sigtermans](https://mischa.sigtermans.me) Menu Close [Thoughts](https://mischa.sigtermans.me/thoughts) [Books](https://mischa.sigtermans.me/books) [Consultancy](https://mischa.sigtermans.me/consultancy) [About](https://mischa.sigtermans.me/about) [← Thoughts](https://mischa.sigtermans.me/thoughts) July 9th, 2026 · AI I stopped running agent teams and moved everything to Solo ========================================================== On a Monday I wrote down that teams beat raw subagents. On Tuesday I moved every subagent to Solo and never made a team again. Here's what changed my mind. On 22 June I wrote a note to myself: always use teams over raw agent spawning, we learned this the hard way. On 23 June I told Claude to run every subagent through Solo instead. Thirteen workers spawned in that one session. I haven't created a team since. Nothing about teams was broken. Something about them was temporary, and I hadn't noticed how much that cost me. What a team is, and where it ends --------------------------------- Claude Code's agent teams let one session spin up teammates that talk to each other. It's good. You describe the work, the teammates split it, and you watch them go. The catch is the container. A team lives inside the session that created it. Close the window and the teammates are gone. Start a new session tomorrow and nothing they learned comes with them. Every teammate is also Claude, which is fine until the job would be better done by something else. So the shape of the work has to fit inside one sitting in front of the terminal. Most of my work doesn't. What Solo does differently -------------------------- [Solo](https://soloterm.com) is the terminal harness I run everything in. Its unit isn't a teammate, it's a process. A process has a project, a name, an id, and a life of its own. That one difference changes four things. **A worker outlives the session.** I can close the laptop. The worker keeps running, and its output is still there tomorrow, in the same project, under the same id. **A worker doesn't have to be Claude.** Anything that runs in a terminal can be a worker. Codex sits next to Claude in my projects now. When I want a second opinion from a different lab on an architecture call, that's a process too, not a prompt. **State lives outside the chat.** Scratchpads and todos belong to the project, not the conversation. A worker writes its report to a scratchpad. The next session reads it without me pasting anything. **Supervision stops being polling.** This is the part I underestimated. The orchestrating session doesn't check on workers. It asks Solo to wake it when a process goes idle, and Solo does. In the Stagent talents refactor in early July, one orchestrator ran 34 of those timers and barely edited a file itself. It briefed, waited, verified, and briefed again. The brief that stopped being ad hoc ----------------------------------- By the first week of July, the way I start a worker had become a template. Every brief has the same parts: - **A name and a Solo project.** 'You're the X worker for Stagent, Solo project 5.' - **A lock.** Two workers in the same repo grab different locks, so they can't touch the same thing. - **Its own test database.** `stagent_tests_` per worker, so a migration in one doesn't break the suite in another. - **A scratchpad to report into**, not the chat. - **Never push.** Prepare the work, leave the recording to me. - **Escalate product questions.** If it's a product or scope call, the worker asks the orchestrator, who asks me. - **Prove it.** Claims get verified against the repo, not accepted on the worker's word. The last two are the ones I keep paying for when I skip them. A worker that decides a product question on its own produces work that looks finished and is wrong. A worker whose 'done' nobody checks produces the correction I've written more than any other: it isn't done. The tiers I use now ------------------- Teams disappeared from my rules. What replaced them is a list of four tiers, and one line above it: use the cheapest one the work survives. 1. **Inline.** One file, one grep, one command. Claude does it in the session. 2. **Native subagents.** Read-only fan-out inside a session: parallel searches, audits, research. They're cheap, they're parallel, and they return an answer instead of a file dump. 3. **Workflow.** Deterministic multi-agent control flow, when the order genuinely matters. Only when I ask for it. 4. **Solo.** Anything that spans sessions, needs supervision, uses non-Claude workers, or should stay out of the chat buffer. Most work is tier 1 or 2. That's the point of writing the list down. Before it existed, every interesting task drifted upward into more machinery than it needed, because more machinery feels like progress. The lesson under it ------------------- The instructions in my repos said teams-first for two more weeks after I'd stopped using teams. On 6 July I switched the feature off entirely and rewrote the rule. That gap is the real lesson. My practice had moved and my written instructions hadn't, so every new session started by reading something that was no longer true. An agent doesn't know which of your documents is stale. It reads the one you gave it and believes you. Teams end when the window closes. Workers don't. And a rule that describes last month's workflow is worse than no rule at all. *thanks for reading* Hi, I'm [Mischa](https://mischa.sigtermans.me/about). I've been *shipping products* and *building ventures* for over a decade. First exit at 25, second at 30. Now Partner & CPO at [Ryde Ventures](https://ryde.ventures), an AI venture studio in Amsterdam. Currently shipping [Stagent](https://stagent.com), [TAP](https://tap.fm) and [Steddle](https://steddle.com). I [write](https://mischa.sigtermans.me/thoughts) about what I learn along the way. [More about me](https://mischa.sigtermans.me/about). Keep reading: [The MCP server that can't see who you are](https://mischa.sigtermans.me/thought/the-mcp-server-that-cant-see-who-you-are). [← Thoughts](https://mischa.sigtermans.me/thoughts) Connect ------- [X](https://x.com/mischamartijn) [LinkedIn](https://linkedin.com/in/mischasigtermans) [GitHub](https://github.com/mischasigtermans) © 2026 Hold My Beer B.V. · [RSS](feed:https://mischa.sigtermans.me/feed)