Spawn Agents

Spawn Agents

Overview

Delegate independent work to subagents with isolated context. By precisely crafting the agent's prompt, you ensure it stays focused and succeeds. Subagents inherit nothing from your session — you construct exactly what they need. This preserves your own context for coordination.

Core principle: One agent per independent problem. Parallel only when truly independent. Over-dispatching is more common than under-dispatching.

The Decision: Dispatch or Not?

Dispatch when ALL of these hold:

• 2+ problems exist that are genuinely independent
• Each can be understood without context from the others
• Agents won't edit the same files or rely on the same in-flight changes
• You'd otherwise lose tokens to context bloat investigating sequentially yourself

DO NOT dispatch when:

• The problems share a domain model, schema, or core utility (the bugs may have one root cause)
• The work is one cohesive feature decomposed into pieces (e.g. "implement auth: signup + login + reset + verify" — all four touch User, middleware, and schema)
• You don't yet know what's broken (exploratory debugging needs full context)
• The task fits in one agent (don't shard for the sake of sharding)

Over-dispatch is the failure mode to fear most. Capable models reach for parallelism reflexively. Resist: if streams share state, agents will conflict and you'll spend more time integrating than you saved dispatching.

Soft cap: ~3-4 agents per message. If you're tempted to spawn 5+ in parallel, the more likely diagnosis is that they aren't truly independent — re-examine the decomposition before dispatching.

The Mechanism (Claude Code)

Parallelism in Claude Code is NOT a special Task() function. It is:

Multiple Agent tool invocations in a single assistant message.

The harness runs them concurrently. Sequential Agent calls across messages run sequentially.

✅ One message containing 3 Agent tool calls → all 3 run in parallel
❌ Three messages, one Agent call each → all 3 run sequentially

When dispatching N parallel agents, emit all N tool calls in one block.

Subagent Dynamics

Spawned subagents auto-load using-kisune, which has <SUBAGENT-STOP> — the subagent skips the bootstrap dance and runs your assignment directly. Implications for your prompt:

• The subagent will NOT auto-invoke other kisune skills unless you tell it to.
• The subagent has NO conversation memory — your prompt must be self-contained.
• If you want the subagent to follow TDD or use completion-validation, say so explicitly.

Prompt Construction

Brief the agent like a smart colleague who walked into the room — they have not seen your conversation, do not know what you've tried, do not know why this matters.

A good dispatch prompt has four parts:

1. Goal — one sentence on what to accomplish
2. Context — relevant file paths, error messages, prior attempts, why this matters
3. Constraints — what NOT to change; scope boundaries
4. Expected output — what the agent should report back, in what shape

Investigate why src/auth/middleware.ts:42 throws "token not found"
on valid sessions.

Context: Started after PR #312 merged the session-cookie refactor.
Failing path: GET /api/me with valid cookie → 401. Repro: tests/integration/auth.test.ts case "valid session returns user".

Constraints: Do NOT modify production code. Identify root cause only.
You may add console.log in tests.

Return: Root cause in 1-2 sentences, the exact line(s) responsible,
and a proposed fix (do not apply it).

When "Independent" Turns Out Wrong

Mid-flight, you may notice agent A and agent B are converging on the same root cause (shared utility, same schema column, same race). When this happens:

1. Stop further work — if either agent has already returned, do not commit those changes blind.
2. Consolidate — open a single new investigation that covers both problems together.
3. Re-dispatch only after the shared root cause is understood; the new agents must not overlap on the now-shared edit surface.

Sunk-cost fallacy: "the agents are already running, let them finish." Resist. Conflicting edits cost more to untangle than to abort.

After Agents Return

For each returned agent:

1. Read the summary critically — agent reports describe intent, not necessarily reality.
2. Verify via diff/state, not the report. Run git status, git diff, read the changed files. (See completion-validation for the discipline gate before claiming "the agent succeeded".)
3. Check for cross-agent conflicts (same file edited twice, same dependency added differently).
4. Run the full test suite once integrated — independent fixes can still combine into a regression.

Partial failure (1 of N fails): Do not silently drop the failed agent's task. Either (a) re-dispatch it with the lessons learned from siblings, (b) take it on yourself if it now needs full-system context, or (c) report the partial state honestly: "2 of 3 fixed, the third needs further investigation". Never claim the whole batch succeeded when one stream failed.

Common Mistakes

When NOT to Spawn

• Scoped lookups already known: use Read / Grep directly, not an agent.
• One-shot tool execution: don't wrap a single Bash call in an agent.
• Short reactive tasks: if the next action is dictated by what you just read, do it yourself.
• Anything where you'd give the agent the same context you already have: dispatching adds latency without isolation benefits.

Integration With Other Kisune Skills

• using-kisune — defines the <SUBAGENT-STOP> semantics that let dispatched agents skip the bootstrap.
• completion-validation — the verification gate after an agent returns. "Agent said success" is not evidence; check the diff.
• brainstorming — if the dispatch decision itself is non-obvious (independent vs not?), brainstorm before dispatching.
• spec-driven-implementation — for Full-mode features, do NOT shard task execution across agents unless tasks are explicitly independent in the plan.