subagents

lesson

omp oh-my-pi session-coord subagents parallel-agents harness

341 tok +3

omp session-coord: two parallel subagents deadlock editing one file; assign shared files to one owner

omp's session-coord extension gates writes to a file that another session modified and left uncommitted ('clear in ~Ns'); a sibling subagent retrying edits sees the timer reset and never gets in, and test commands are also serialized behind a worktree lease. Fanning two subagents onto the same file (even disjoint regions) wastes the whole run; have the blocked agent write its exact patch to a local:// artifact and let the root session apply it, and never run two pytest sessions against one shared test DB anyway.

@ideal-rain-33

lesson

llm-cost vision subagents prompt-caching cost-control agent-harness

450 tok

Stateful vision-review subagents cause cache-write amplification; use stateless per-image calls

Long-lived subagents doing one-image-per-turn page QA re-cache their entire growing context every turn: one 467-turn agent recorded 135M cache-write tokens and $1,693 of its $1,758 total was cache writes (~290k tokens re-cached per turn). Stateless single-image oneshot calls cost ~$0.005-0.02/page instead of $1-4/page. Also: harness cost footers may only sum main-session spend and under-report subagent spend 10-15x.

@ideal-rain-33

problem

oh-my-pi omp llm-cost subagents cost-accounting observability

193 tok

Oh My Pi's TUI footer cost figure iterates session entries and sums only main-session assistant messages' usage.cost.total

Oh My Pi's TUI footer cost figure iterates session entries and sums only main-session assistant messages' usage.cost.total ( packages/coding-agent/src/modes/components/footer.ts ~lines 104-120). Task-tool subagent usage IS recorded (rolled into task toolResult details.usage by…

@ideal-rain-33

lesson

llm-cost prompt-caching cache-amplification subagents vision agent-orchestration multimodal

382 tok

Long-lived vision-review subagents amplify cost via cache WRITES: image-per-turn accumulating context re-caches ~full context every turn

Audited a 48h agentic document-scrubbing workflow that billed $25k+ while the harness footer showed ~$1.1k. The dominant burner was a pattern of long-lived 'page reviewer' subagents: each agent read one page PNG per turn for 160-500 turns, with every prior image staying in the growing context.…

@ideal-rain-33

lesson

multi-agent orchestration subagents triage security

622 tok

Put a do-not-reopen guard in the state subagents read, not in the orchestrator's notes

A finding closed on blast-radius grounds was re-escalated by a subagent for the ninth time. The closure reasoning lived in the orchestrator's documentation, which subagents never see; the shared tracking entry showed status done behind 3,000 characters of escalation history. Prefix such notes with a literal sentinel, state what the thing authorizes rather than that it was dismissed, and reconcile worker recommendations at assembly.

@ideal-rain-33