← sources

Our planner agent kept doing the work itself, so now it can't write a single file during its turn

source: reddit (reddit:AI_Agents)published 2026-10-12 19:52ingested 2026-10-13 02:36

retained text

Atomic Agent (open source, MIT) has a mode called Fusion: one model plans, splits the job into parts and reviews; worker models build the parts, several at once. Usually a cloud model plans and local models do the work, so only the planner spends API tokens. The direction can be flipped.

The rule that took us longest to get right is what the planner is allowed to touch. The answer we ended up with is nothing. Read, delegate, reply.

**How we got there**

The first gate was soft. It closed writes for the planner until one delegation succeeded, then opened. Two things broke it.

First, `fusion.delegate` returned `ok` even when every worker had failed. The planner saw ok, the gate opened, and the planner wrote 15 files and ran 5 commands itself. The run "worked". Fusion didn't happen.

Second, we let workers hand a task back with `needs_orchestrator` when it was too big or unclear, and that opened an exception for the planner. 13 writes went through that hole. In that run, 4 of 6 tasks came back that way.

Earlier still there was a version where the gate was dead code: the step executor only forwarded plan-mode, so the check never ran. A driven test wrote all four files from the planner with zero refusals. The gate was live and unit-tested, it just never got consulted.

**How it works now**

For the whole turn the planner can't mutate anything. MCP tools count as mutating no matter what their hints say, because they're third-party code. Workers write to files, the planner reads them, runs read-only checks, and re-delegates.

The refusal is a message in the tool result, not a hidden tool. If we removed tools from the planner's catalog, the prompt prefix would change and a local planner would lose its KV cache every time the rule kicked in. So the tool stays visible and calling it gets a short "you plan, workers write".

**The irony**

The guardrail that hurt most in our measurements was this one. A local planner found a real bug, diagnosed it correctly, and then spent 22 minutes writing the fix into the text of its refusal, because it wasn't allowed to write the file. We kept the rule anyway. A planner that writes is how Fusion quietly stopped being Fusion twice. A planner that's slow to hand off at least shows up in the trace, and the fix there is a better brief to a worker.

**Bringing your setup over**

The other question we get from people running other agents: what moves over. There's an importer for six agents. Every importer has `--dry-run` that writes nothing, and API keys only move with an explicit `--migrate-secrets`.

|From|What it imports|Keys with --migrate-secrets|
|:-|:-|:-|
|Claude Code|skills, memory, MCP servers, sessions|ANTHROPIC\_API\_KEY|
|Codex|skills, AGENTS.md instructions, sessions|OPENAI\_API\_KEY (empty on ChatGPT login)|
|Hermes|sessions, cron jobs|OpenRouter / AI/ML API keys|
|OpenClaw|sessions, cron jobs|no|
|Pi|skills, sessions|no, out of scope|
|Oh-My-Pi|skills, MCP servers, sessions|no, out of scope|

Memory from Hermes and OpenClaw doesn't come over. The stores are different and there's no exchange format, so I'd rather say so than have you find an empty memory tab.

It runs as a terminal agent and, since this week, as a desktop app (Mac, Windows, Linux). Approvals in the desktop are Allow once, Deny or Abort run; to be asked less you switch the mode in the composer.

Repo link in the comments.

If you run planner/worker setups: do you let the planner write at all, or did you also end up at zero?

document fetched 2026-10-13 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1x4c1vu/our_planner_agent_kept_doing_the_work_itself_so