agents

Agents

Command-driven skill for building, diagnosing, and running agent systems (Cursor/Codex/Claude agents, Task/subagent trees, Trigger/durable workflows, custom orchestrators). Parse the first token as a command when it matches the table; otherwise map clear intent.

Operating contract

Shared by every DecisionNerd/dev-skills skill. Canonical copy: handbook/concepts/14-operating-contract.md.

• Define done first. Before using tools, write the finish line in one or two lines: acceptance criteria (existing issue AC, BDD scenarios, tests, or contract; otherwise propose them and where they should live) and the evidence that will prove them. Re-check it before reporting done. Never report done on unverified work.
• Requested scope sets the finish line. A question ("is it ready?", "why is it broken?", "what next?") ends with the answer and a Next: line naming the exact next invoke. An outcome request ("fix", "finish", "land", "#42") continues through the chain (diagnose → fix → test → check-readiness → merge-it) until the outcome or a real blocker. Do not end a turn with "Do you want me to…?" for in-scope, in-repo work.
• Stop only for real blockers. Stop and ask only when you cannot continue without the user, or before: deleting data or unmerged work, force-push or history rewrite, changing anything outside this repository (GitHub objects, deployments, live data, production or paid resources, external services), or leaving the requested scope, unless the user's request already named that exact action. Keep the harness's permission prompts for risky commands. Otherwise keep going and put status notes in the same message as the next action.
• Ask well, once. For a genuine question use the harness's structured question tool when it has one (Claude Code: AskUserQuestion; Codex: request_user_input when the current mode supports it; T3 Code: the question tool) with concrete options; otherwise plain text with numbered options. Treat the answer as settled; do not re-open earlier verdicts, plans, or answers unless asked.
• Fan out when work is parallel. Use subagents for independent reads (repo survey, evidence gathering, per-option research, per-area audits) and independent verification (a reviewer that did not write the change). Writes stay single-owner per path set and sequential. Brief every child with goal, done-when, constraints, must-not, and return shape; verify each result before consolidating. Use Claude Code's Workflow tool only when orchestrating many subagents truly needs it; it is expensive. T3 Code: its delegated subagents — with optional worktree isolation for independent verification — cover this without the Workflow tool.
• Pick the model tier per child; defer to routing config. If the harness or user config already routes subagents, follow it and do not pass a model. Otherwise pick a tier per child: mechanical search or inventory → fast/cheap; implementation and evidence gathering → mid; planning, review, adversarial verification → top. Tier names differ per harness (Claude Code: haiku / sonnet / opus or fable; T3 Code: inherit by default or pick from its live provider/model catalog). Set the choice with the harness's spawn model param or agent frontmatter model: (per-harness details: skills/agents/references/sub.md). Where the harness cannot choose, children inherit or the harness picks; say which in the status note.
• Keep a checklist on long runs. For more than about five steps or work that crosses skills, keep TASKS.md at the repo root and tick items as they finish. Do not commit it unless the repo already tracks one.
• Close every run with three headings. Blocked on me (the one genuine question or blocker, else "none"); Changed (files, commits, GitHub objects, deploys, else "nothing"); Found (evidence, verdict, and Next: <exact invoke>).

Commands

Aliases: subagents → sub; drain / stop / kill-loop → prefer slap when the situation is emergency thrash.

Target: path to agent config / skill / workflow, run ID, PR/issue, or free-text symptom (“it keeps rewriting the same file”).

Routing

1. No argument: ask intent via the harness's structured question tool (Claude Code: AskUserQuestion); options: slap (something on fire), analyze, optimize, design, sub, help.
2. First word is a command: run it; remainder is target/context.
3. Clear intent (“emergency stop this agent”, “design a code-review agent”, “use subagents to explore”): map and proceed.
4. Active runaway loop / dumb workflow burning tokens: default to `slap` even if the user only said “fix the agents”.

Mutating production workflows, killing remote runs, or deleting agent state requires explicit approval unless the user already ordered that exact action. Prefer pause/drain over hard kill when both are possible.

slap — emergency fix + safe drain

Use when an agent is thrashing, looping, spamming tools, rewriting the same files, ignoring instructions, or a durable/workflow run is producing junk. Follow references/slap.md.

Goals, in order:

1. Stabilize — stop further damage (pause run, cancel queued steps, freeze writes).
2. Drain — let in-flight safe work finish or abort cleanly; no orphan locks/partial deploys.
3. Contain — revert or quarantine bad agent edits if they landed.
4. Slap — name the stupid behavior in one blunt sentence + the smallest durable fix (prompt/guardrail/tool allowlist/max-steps).
5. Handoff — analyze for root cause, optimize/design for lasting change, fix-it if product code broke.

Do not start a new mega-agent to “fix” the runaway. Parent stays in charge; subagents only for narrow read-only recon if needed.

analyze

Read-only. Produce:

• Surface — what agent(s), which harness (Cursor Task, Codex, Trigger, custom), entry prompts/skills.
• Behavior — observed loops, tool misuse, missing context, contradictory instructions.
• Evidence — logs/traces/transcripts/run IDs; cite concrete steps.
• Failure modes — ranked: loop, context rot, tool hallucination, scope creep, unsafe writes.
• Verdict — keep / redesign / slap-then-optimize.
• Next — exact follow-up (agents optimize …, agents design …, agents slap …).

Details: references/analyze.md.

optimize

Only after you know what is bad (analyze first unless the user already provided a sharp diagnosis). Propose a small numbered plan (prompt splits, tool narrowing, caching, evals, step limits, model routing). Implement in-repo prompt and config changes; ask only before production-workflow or paid changes. Prefer measurable wins (fewer steps, fewer retries, clearer success criteria).

Details: references/optimize.md.

design

Architecture before code. Output a short design: goal, inputs/outputs, tools (allow/deny), memory, success criteria, failure/escalation, whether to use sub fan-out, eval plan. Scaffolding is a separate step — end with Next: scaffold <name> rather than asking. Prefer one sharp agent over a swarm.

Details: references/design.md.

sub / sub-agents

Parent agent orchestrates; children do bounded work. Follow references/sub.md.

Rules:

• Write a crisp parent brief: goal, constraints, done-when, tools allowed, what not to do.
• Prefer parallel explore/research children; keep write/mutate work sequential or single-owned.
• Never give two subagents overlapping write ownership of the same files.
• Verify each child’s result before consolidating; do not rubber-stamp contradictory child plans.
• Aggregate verified results; parent decides.
• If children thrash → `slap`, don’t spawn more.

Output

Close every run with:

Blocked on me: (production/paid change awaiting approval, or "none")
Changed: (files, config changes, GitHub objects, else "nothing")
Found: (analysis verdict, plan, or design). Next: `<exact follow-up invoke>`.

Related skills

• Product/backend bugs: diagnose-bug / troubleshoot-app / fix-it (do not use agents as the primary cut for deterministic compute or UI data-plane bugs)
• Tests / evals as success criteria: test-it (regime C)
• Traces/scores in prod: observe-it (Langfuse house default for LLM)
• Docs for agents/humans: document-it (surgical)
• Options before design: research-it
• Repo situational awareness: recon

Grounding

This skill’s TTPs are grounded in current engineering baselines (DORA, GitHub Docs, Fowler/Beck, Google SRE & SWE book, OpenTelemetry, OWASP LLM / NIST AI RMF, Diátaxis — see handbook sources.md).

Anthropic: keep agents simple and evaluable. OWASP LLM: bound tools/agency; NIST AI RMF: measure before optimize. Use slap to drain runaway loops (excessive agency). Parallel explore / sequential mutate for sub-agents. This is quality regime C — design for traces + evals (Langfuse house default), not unit-test theater (handbook/concepts/11-quality-regimes.md).

Handbook card: handbook/practices/design-agents.md.