<!– story: e45s11 –>

HARD GATE — HARD GATE — Agent work must be parallelizable and have explicit synchronization points. Do NOT dispatch work that has hidden dependencies between agents.

Run multiple subagents in parallel on independent tasks. Use when tasks are genuinely decoupled — no agent needs the output of another to start.

Distinct from delegate-task: This skill maximizes throughput via concurrency. There is no sequential review gate between tasks. Use delegate-task instead when a single task needs careful two-stage oversight before proceeding.

• Tasks that can run simultaneously without shared state
• Large plans that can be broken into parallel workstreams
• Exploration: gather information from multiple parts of the codebase at once

• Task B depends on Task A’s output
• You need to review Task A before Task B can start safely
• The tasks share a file and concurrent edits would conflict

Before dispatching, verify each task pair is truly independent:

• No shared files being written
• No shared state (DB migrations, config files)
• No ordering dependency between outcomes

If any two tasks conflict, sequence them with delegate-task or execute-plan instead.

Map effort: frontmatter and story risk: to prompt depth — do not send minimal_decisive agents a full_maturity brief.

Record depth: <tier> in the Agent tool description when dispatching.

Before writing briefs, read specs/state.yaml if it exists — each agent gets only the decisions relevant to its task, nothing else.

Every inter-agent message uses a typed envelope — no freeform prose between waves:

Example task_brief (each agent starts cold — brief size directly controls token cost and hallucination risk):

type: task_brief
task_id: agent-1
goal: [one sentence — what success looks like]
in_scope: [explicit file or module list]
out_of_bounds: [what NOT to touch]
verify: [runnable command]
prior_decisions: [relevant entries from specs/state.yaml — omit if none]

Emit checkpoint comments when an agent is slow or blocked — one line, no stack traces. Parent reads checkpoints before spawning follow-ups.

Do not include the full conversation, full file contents, or decisions unrelated to this agent’s task.

Track consecutive failures per task_id. On the 3rd consecutive result.exit: fail for the same task, emit type: circuit_open and stop dispatching that task — escalate to user with the three failure summaries. Reset counter on any pass.

After each wave completes:

1. Dispatch — run parallel agents with typed task_brief envelopes.
2. Evaluate — read result messages; list gaps vs goal; honor open circuits.
3. Refine — tighten briefs or spawn follow-up agents (max 3 cycles total).

Stop when gaps empty, circuit opens, or cycle 3 reached — escalate to user. If agents go silent without returning, invoke diagnose-stall before spawning another wave.

Spawn all agents in a single message using multiple Agent tool calls. Each agent gets its own complete brief.

Agent 1: brief for task A
Agent 2: brief for task B
Agent 3: brief for task C

When all agents return: review each result, run verify commands, check diffs for scope violations.

Merge accepted results. Resolve conflicts manually; note in summary.

Report: which tasks succeeded, which need revision, overall verify status.

→ verify: grep -q 'circuit_open' skills/dispatch-agents/SKILL.md && grep -q 'task_brief' skills/dispatch-agents/SKILL.md && echo OK || echo FAIL