Subagent transparency — show what they do, don't silent-pass-through

When spawning a subagent, show what it will do + return format + verification plan. Summarize subagent output back to user before acting on it.

Subagent transparency

The rule

When spawning a subagent, the user MUST see:

1. What the subagent will do — one-line goal in plain language
2. Return format — terse summary paragraph? structured table? file path?
3. Verification plan — how the main agent will check the subagent’s output

And when the subagent returns:

4. Summarized output — main agent must summarize back to the user before acting on the subagent’s return, unless the output is a single well-scoped artifact (e.g. a decision file path).

Anti-patterns

• ❌ Silent pass-through: subagent returns 300-line dump → main agent acts on it without summarizing
• ❌ Subagent output triggers a series of tool calls the user never approved
• ❌ 5 parallel subagents fire but only 3 tasks visible in TaskList — the other 2 are hidden
• ❌ Subagent hallucinated output → main agent trusts it without verification → cascading errors (the “Lemonade insurance” pattern at agent-of-agent level)

The transparency contract

For every Agent tool call:

Task #N: <one-line what>
Agent goal: <what success looks like>
Return: <expected shape>
Verify: <how I'll check the output>

Fire the Agent call AFTER this contract is on-screen. When the subagent completes, first-line of your response summarizes what came back before any acting-on-it happens.

Special case: background subagents

Background subagents (run_in_background: true) are especially prone to silent-pass-through because the completion notification arrives mid-work-on-something-else. When a background subagent completes:

1. Announce completion + one-line what it did
2. Show the returned artifact/summary
3. WAIT for the user’s next input before acting on the output — unless the user’s original directive was clearly “and continue from what it returns”

When transparency can be tighter

• Trivial subagent (single grep across N files) — skip the full contract; a one-liner is fine
• Subagent whose output is a specific file path — main agent reading the file counts as visible action
• Fan-out of ≥5 subagents in one message — a single group announcement covers all instead of per-agent contracts

Anti-anti-pattern (when transparency hurts)

Don’t over-narrate:

• ❌ “I’m now going to spawn a subagent to do X. First I’ll define the goal. Then I’ll write the prompt. Then I’ll…”
• ✅ “Spawning: . Return: . Verify: . Firing now.”

Cross-refs

• delegate-to-subagents-by-default — WHEN to delegate
• ticketing-primitive — every subagent spawn = a task
• draft-not-send — subagents that touch external state = user-approved before send
• own-memory-rent-intelligence — the broader principle
• Source: Nate B Jones 2026-06 agentic-loop visibility + Nikita/Lemonade insurance-story lesson

Edit on GitHub · Back to index