Repository navigation
[idea] Lean sub-agent payload + 250k autocompact default — sub-agents load ~145–215k before first write #167
Description
The idea: the cook-off's cost gap was context per call, not number of calls (29.5 vs 28.7 calls per row). Two causes:
- The lead ran with autocompact accidentally set to 1M.
- Sub-agents load the whole always-on set plus the full build plan and BRD (~145–215k), then never compact. 34 of 38 sub-agents peaked at 250–580k.
Proposals:
- Role-specific stubs that name 2–4 skills.
- Give each agent its per-row spec slice, not the whole spec.
- A 250k autocompact default, or a preflight warning, written by
init-project.sh. - A cap on how many rows are bundled per agent.
- Per-role mean context per call recorded in the time log.
Field evidence: contrib/inbox/2026-09-29-subagent-context-cost.md (#165). Re-estimated at 250k autocompact, the build cost drops from $264 to ~$205.
Done = on the next comparable run, calls per row are unchanged, mean context per call is about halved, and cost per row is at or below the MCP arm's.
Rough size: a cluster (agent-roles.md, agents/ stubs, init-project.sh).
Activity
MendixMau commentedon Oct 12, 2026
Info for research (no PR requested). Retro 2026-10-12, measured fixed prefix.
In a 89-agent run every subagent started at 47-48 k tokens before reading its brief (first-turn cache read plus cache creation: 17.6 k + 29.8 k in a sample). The always-on parts, in characters: skill listing 66 k, instruction chain 76 k, deferred tools / agent list / MCP notes / session context about 35 k. 48 k x 4,061 subagent turns is about 195 M cache-read tokens, 40% of subagent volume; the controller carries the same prefix on 1,204 turns. Subagents compacted 20 times, reaching about 215 k context; average context for implementers was 125-150 k. Estimated saving from 12-15 k fewer prefix tokens per turn: 6-8% of weighted units. Details and the rest of the token analysis are in the new issue "Token anatomy of a 20-task, 89-agent run".
🤖 Generated with Claude Code
Merge #323: inbox promotions — exec refusal table, UI fix rounds, mea…
MendixMau commentedon Oct 19, 2026
Status after 1.0 (v2026.10.19): init-project.sh now recommends
/autocompact 250kand the role stubs are trimmed (#329, #223, #385). Still open: a per-row spec slice for sub-agents and per-role context figures from a measured run.MendixMau commentedon Oct 20, 2026
Closing as completed (maintainer decision, 2026-10-20). The substance shipped, verified on master:
The role stubs still carry 55–85 trigger rows rather than 2–4 skills, but stub truncation (#385) and on-demand routing keep the always-loaded cost measured. Further trimming is tracked by #166 (context diet) and the release A/B.
Generated by Claude Code