Blog

3 min readrelease

The delegation round: every spawn now says why

Unown 0.0.46: the why field, the delegation audit ledger, four specialist agents, and a measured 76→79 with the runaway-scout shape dead — receipts inside.

  • #subagents
  • #observability

The most common complaint about subagents in the field is not that they exist — it is that they are unaccountable. Claude Code's own tracker has a measured 2.41B-token run where 71% of subagent tokens burned after each agent's 80th step, and a user asking the agent to state its premises before delegating so I can correct early. GitHub shipped a trajectory-analysis-driven delegation fix and measured tool failures down 23%. The field knows delegation quality is the gap; almost nobody instruments it.

This round, Unown did. Every spawn now carries its reason.

What shipped

The why field. Every task call takes one line — why you are delegating this instead of doing it inline, and what you expect back. It is not decoration: it rides the tool result, the async ack, the fleet delivery, task_control status and /tasks, it lands in the fleet journal, and it reaches a new delegations table the dashboard renders. In the after-benchmark the model stated it unprompted on 12 of 13 spawns: "Parallelism: user wants the count in the background while answering a question inline." That is the premises-before-delegation ask, answered.

The audit ledger. Both lanes' terminal spawns become records — why, turns, tokens, tool calls, wall clock, outcome — in a per-session ledger (/delegation card + footer chip), a cross-session engine aggregate (unown delegations, the dashboard's new Delegation screen), and the durable journal. Verdicts are receipts-only: a blocked or timed-out run is not worth it by its own status, a clean run reports facts, and an expensive one (≥100k tokens) is flagged for audit — never judged on vibes.

The runaway is dead. The baseline caught a scout burning 104 turns, 596k tokens and 111 tool calls on one question, killed only by the 300 s wall clock — the exact shape Claude Code issue #106025 measured in production. The read-only agents now ship turn caps (24/24/40), and in the after-run that lane finished in 43 s with all three scouts clean. Child token spend fell 62%; total workload spend fell 54%.

Four specialist agents. debug (read-only, hypothesis-driven root-cause diagnosis), test (baseline → write → re-run, completion guarded), security-reviewer (severity tiers with exploit preconditions and an INSUFFICIENT CONTEXT verdict), documenter (verified file:line references, never invents APIs) — plus two project agents that audit Unown itself: delegation-auditor and benchmark-curator.

The receipts

The round ran a ten-axis benchmark on the live agent, before and after, with every number receipt-backed (the full delta table):

Axis 0.0.45 0.0.46
User explainability 56 92
Progress visibility 55 100
Delegation quality 86 84
Overall 76 79

The benchmark also caught three more failures during the round itself — the print envelope shipping the interim answer on async deliveries, the benchmark's own load saturating the dream trigger, and a benchmark session mutating the working tree against the anti-tamper rule — each reproduced, root-caused, fixed or honestly recorded in the F26–F38 ledger. One new behavior is recorded, not hidden: the model sometimes over-retries a delegation (five reviewer spawns for one review ask). That is the top candidate for the next round, and it is exactly what the delegation-auditor polices in production.

The honest limitations stand: n=1 per lane on a metered upstream, a keyword-heuristic handoff score, author-scored rubric axes. The receipts are committed so every number can be hand-audited.

Why it matters

Delegation is the leverage move in agentic coding — a fresh context for the work that would drown the parent's window. But leverage without an audit trail is a bet you cannot price. With why on every spawn, a ledger that answers what did it cost and how did it end, turn caps that kill the runaway shape, and specialists with contracts instead of personas, delegation becomes a decision you can inspect — during the run, after the run, and a month later in unown delegations.

Spotted something wrong or worth adding? The record is the repo — open an issue.

All posts