Agentcraft

---
name: agentcraft
description: Work ON agents and multi-agent systems. Use when authoring an agent, composing a team, choosing a coordination pattern, hardening an unattended run, diagnosing agent memory, running a lightweight work loop, or checking whether a scheduled automation still produces output. Skill craft goes to skillsmith.
metadata:
  author: Zain Haseeb (Quoxient)
  version: 1.0.0
  changed: "2026-07-30"
  verified: "2026-07-30"
  category: meta
  tags: [agents, orchestration, multi-agent, operations]
compatibility: Harness-agnostic. Attest mode needs a shell and Python 3.10+. Cross-harness notes assume Claude Code, Codex, Hermes, OpenClaw, or cmux; the design guidance is portable to any of them.
---

# Agentcraft

The workbench for **agents**, the way skillsmith is the workbench for **skills**.

Skillsmith asks "is this tool built well?" Agentcraft asks "is this crew designed well, and will it still be working tomorrow when nobody is watching?" Different artifacts, different failure modes.

The boundary, so nothing is ambiguous:

| Concern | Owner |
|---|---|
| The SKILL.md artifact — craft, audit, portability, evals | **skillsmith** |
| Agents, teams, coordination, unattended operation, agent memory | **agentcraft** (this skill) |
| Heavyweight spec-driven production of a finished artifact | **praxiz** |
| Prompt construction for a specific model | **prompt-master** |

If the request is "make this skill better," stop and invoke skillsmith. This skill will happily discuss skills and produce nothing useful, because it has no skill-craft references.

## Modes

Pick the mode, then load only its reference file.

| Mode | When | Reference |
|---|---|---|
| **Author** | Design a new agent — persona, tools, model tier, frontmatter | `references/teams-and-patterns.md` |
| **Compose** | Build a team and pick a coordination pattern | `references/teams-and-patterns.md` |
| **Harden** | An orchestration exists but is fragile, expensive, or silently failing | `references/production-hardening.md` |
| **Loop** | Run work through the lightweight loop (not spec-driven, not Praxiz) | `references/work-loop.md` |
| **Attest** | "Did that automation actually run and produce anything?" | `references/attestation.md` |
| **Diagnose** | An agent keeps misbehaving, or memory changes are not sticking | `references/memory-hygiene.md` |
| **Port** | Move an agent or crew between harnesses, or decide where it should live | `references/harness-topology.md` |

### Dispatch signals

- "I keep writing the same subagent prompt" → **Author** (that prompt wants to be a definition)
- "Should this be one agent or three?" → **Compose**
- "This burns too many tokens" / "an agent stalled and the run hung" / "it failed and I found out days later" → **Harden**
- "Just help me get this done" with no spec appetite → **Loop**
- "Is my cron/launchd/Hermes job still working?" → **Attest**
- "I deleted the memory and it still does this" → **Diagnose** (this is the signature behavior-layer symptom)
- "Should this run in Claude Code or Hermes?" → **Port**

Chains worth knowing:
- New crew: **Compose** → **Author** each member → **Harden** before it runs unattended
- Inherited or drifting crew: **Diagnose** → **Harden** → **Attest** on a schedule
- Recurring manual work: **Loop** once by hand → **Author** the agent → schedule → **Attest**

## The governing principle

**Ad-hoc agent prompts are unbuilt agents.** The signal that you need a definition is repetition: the third time you write substantially the same subagent prompt, you are maintaining an agent with no file, no version, and no way to improve it except remembering what you did last time.

The same logic runs one level up. A crew you assemble by hand each time is an unbuilt team. A check you run manually is an unbuilt attestation.

## Design rules that apply in every mode

1. **Tools follow intent.** Advisory agents get read-only tools. An agent that should author its own findings gets `Write`. The most common mis-scope is a reviewer that cannot write its own report.
2. **Least privilege, and mean it.** Tool sprawl is not free — it widens the blast radius of a bad run and makes the agent's job ambiguous.
3. **One lead per team, and it must be the strongest expert for the team's *primary* workflow** — not the most senior-sounding persona.
4. **Match the pattern to the dependency structure, not to the vibe.** A sequential team doing independent work wastes wall-clock; a parallel team with real dependencies produces garbage. This is the most expensive design mistake available.
5. **Unattended means hardened.** Anything running without a human watching needs, at minimum, self-reporting health checks and an escalation path. Everything else in `production-hardening.md` is optional; those two are not.
6. **Silence is a feature.** An agent that reports every cycle drowns the signal you built it to surface. Budget its silence deliberately.
7. **Never claim coverage you did not get.** If a wave ran with three of five agents, the output says so. Degraded results reported as complete are worse than a failed run, because they end the investigation.
8. **Model tier is a design decision, not a default.** Match it to judgment required, not to task importance. Bulk mechanical work on an expensive model is the most common waste in a fleet.
9. **Delegate execution, keep judgment.** On a frontier model, hand mechanical and well-specified work to cheaper subagents and reserve the orchestrating context for synthesis and decisions.
10. **Cross-harness memory does not propagate.** Each harness has its own retrieval layer. A fix applied in one place has not been applied everywhere — see `references/memory-hygiene.md`.

## What this skill does not do

- **It does not write skills.** → skillsmith
- **It does not run the spec-driven production loop.** → praxiz, when the work genuinely warrants interview → research → spec → gates. Agentcraft's Loop mode is the lighter path for when it does not.
- **It does not pick models for a prompt.** → which-ai-model
- **It does not manage skill distribution or presets.** → skillz

## Output conventions

- When authoring an agent or team, **write the actual file**, do not describe one. An agent definition that exists only in a chat message is the problem this skill was built to solve.
- When hardening, name the specific failure the change prevents. "Added a silence budget" is not a finding; "this agent emitted 40 status updates over a 6-hour run and none carried new information" is.
- When attesting, report **last real output**, never "job is registered." A scheduler that fires a broken job on time is a scheduler reporting success.
- Always end with what the operator does next.