Fleet of One

I barely write the code anymore. I orchestrate the agents that do. Notes on what breaks, and what fixed it.

[Fleet of One @fleetofone]

@carlosazaustre genial que se pongan de acuerdo con el nombre del archivo, cada herramienta debería saber qué abrir. pero al final las instrucciones cada modelo las va a interpretar a su manera. si se logra un mínimo común, como dónde se encuentra cada cosa, al menos es un avance.

[Fleet of One @fleetofone]

@dexhorthy took me a while to get there. the steps my agent kept skipping in a skill stopped getting skipped once i moved them into a hook. what i leave in markdown is the stuff that needs judgment.

[Fleet of One @fleetofone]

@thorstenball same trade here. sure, i lose time now and then fixing things i wouldn't have gotten wrong by hand. i still come out ahead, mostly because i couldn't run several projects at once before.

[Fleet of One @fleetofone]

@ddonprogramming agree on re-entry. i write mine down when i make the decision rather than when i switch, a short note in the repo with the reason. coming back is mostly reading that. and yeah, it's the first thing i open when a run goes wrong.

[Fleet of One @fleetofone]

Working on several projects at once with agents was never my problem. Making several hard decisions at once was.

On a normal day I have several projects open, each at a different stage: one I keep fixing with what I learn, one mid-migration, one closing out a plan and checking it, and so on.

What makes it hold is that each plan runs on its own in an isolated worktree, with its own environment, database and e2e setup. While one runs, I switch, make a decision, switch again. The waiting time gets used instead of spent watching the agent.

Where "stop multitasking with agents" is right is planning. Keeping design discussions open in several projects breaks my focus. So one complex thing at a time: the project that needs decisions gets my head, and the rest only moves on work that was already agreed.

[Fleet of One @fleetofone]

lots of people today saying dev work turned into endless prompting. i had that for a while, and i think it came from keeping every decision in my head.

what fixed it for me was writing each decision down once, in the repo, where the agent reads it at the start of every session. i also trimmed what each agent can load and touch.

these days i still re-prompt, mostly when i change my mind about something.

sure, a run still goes sideways now and then and eats an afternoon. but i don't retype context anymore, and i think that was most of the endless part tbh.

[Fleet of One @fleetofone]

@johnlindquist a benchmark like that would mostly tell you which model to chase next. i'd rather go deep on one. learning how it reads my instructions has done more for me than switching, and the model you already know usually gets there too, with more work.

[Fleet of One @fleetofone]

@GeoffreyHuntley agree it's the plan and not the model. the lock-in is what you pay for the subsidy, and it only stings if you wanted a different harness in the first place.

[Fleet of One @fleetofone]

@alexgetmancom compaction keeps whatever the session picked up, not what you would have picked. starting a fresh session with a short brief gets you the same cheap context without the summary in the middle.

[Fleet of One @fleetofone]

@jediahkatz @bot same lesson on my side with claude code subagents. the cost wasn't how many i spawned, it was what each one loaded before doing any work. trimming tools and mcp servers per subagent saved me more than spawning fewer.

[Fleet of One @fleetofone]

@bcherny managing sessions by hand is a big part of my day in the cli, and i'd happily give that part up. the other half is what i care about: a project that remembers how i work matters more to me than the parallel threads.

[Fleet of One @fleetofone]

@omarsar0 mcp for tools, sure. passing context between agents is the part i'd separate: a tool call is a contract, a handoff is a decision about what to drop.

[Fleet of One @fleetofone]

@claudeai shared memory is the part i'd watch. i audited claude code's project memory across my repos: of hundreds of entries, fewer than ten belonged there. most of the rest were duplicates or work already closed.

[Fleet of One @fleetofone]

this week i ran a project from cowork and dropped into claude code whenever something had to execute. merging chat into cowork makes sense. the line i kept crossing was the other one: cowork to code.

[Claude @claudeai]

Claude Cowork and chat are merging into one Claude.

Ask a quick question or hand over a report, and Claude takes it from there, even after you close your laptop. If something's unclear, Claude asks—you keep the final say.

Rolling out to Pro and Max over the next few weeks.

[Fleet of One @fleetofone]

@claudeai give it a year and claude desktop is just claude code with a nicer window.

[Fleet of One @fleetofone]

@omarsar0 coding is where they work best for me. the main session writes a brief per task, a subagent implements it, the main one reviews and commits. that's how 30 tasks of one plan got done without it stopping on me.

[Fleet of One @fleetofone]

@thekitze what fixed it for me was not asking the main session for all four. it hands each one to a subagent and reviews what comes back. a subagent can't stop to check in with you, so the list gets done.

[Fleet of One @fleetofone]

@sydneyrunkle and 2 is where most of the savings are. switching sonnet to haiku saved me a little. deciding which tools, skills and MCP servers each subagent loads saved me far more.

[Fleet of One @fleetofone]

@ddonprogramming Yes, but not by design: Claude Code logs every turn with its tool calls and results. The catch is divergence. Change the handoff depth and turn one branches, so what's stored after it is a run that no longer exists.

[Fleet of One @fleetofone]

The context-window knee for a coding agent session is not 150k tokens. It is 200k, and it holds whether re-entry costs you 43k or 95k tokens.

I replayed 2,104 real Opus 5 sessions under a "hand off at a fixed depth" policy, each one using its own observed context growth. Marginal tokens saved per extra handoff:

300k: 34.7M
250k: 23.1M
200k: 19.5M
150k: 10.3M
100k: 3.0M

The biggest proportional drop is between 200k and 150k. Below that you pay re-entry (about 30k tokens for turn 1 of a fresh session, plus re-orientation) to save less and less.

Total input saved cutting at 200k: 45.6%. With a pessimistic re-entry cost, 60k floor plus 35k re-orientation, it is still 37.7%.

150k is where /context in Claude Code starts flagging heavy consumers. That is a reporting threshold, not a performance one.

First pass was 638 sessions. Tripling the corpus moved the 300k figure 29% and the 200k figure 10%. The knee stayed put.

[Fleet of One @fleetofone]

@dexhorthy the intern analogy has a second half: you don't only keep them off the migration, you hand them the conventions. every line in mine is there because i said it once and had to say it again.