Portable agent workflow design

Design workflows that stay portable.

Meta Harness turns a project goal into a domain-specific workflow, the specialist skills that support it, and the handoff artifacts that keep the work inspectable.

Current release: v0.8.4 · Runtime-neutral by default · Open source on GitHub

Start with the shortest useful path

Install the shared skill into a project with the meta-harness CLI, then ask for the smallest reusable workflow that meets your need:

meta-harness install \
  --scope project \
  --target /path/to/repo \
  --agent generic \
  --non-interactive

If you are using a source checkout without the CLI on PATH, run python scripts/install_harness.py with the same subcommand and options. The checkout command uses the same modern planner; the legacy --layout form is for migration only.

The installer creates .agents/skills/harness/ and leaves the target repository’s AGENTS.md, README.md, and documentation under the target project’s ownership.

1. Describe the goal

State the domain, the final deliverable, and the quality bar before choosing a team shape.

Browse prompt patterns →

2. Choose the shape

Use one of six coordination patterns only when specialization, ordering, or review provides real value.

Compare patterns →

3. Preserve the handoff

Keep durable specs, role briefs, and intermediate artifacts in predictable repository paths.

See output specs →

What Harness gives you

6
execution phases after the Phase 0 audit
6
coordination patterns
1
portable canonical skill tree
0
required runtime-specific agents

The core design is intentionally small:

Phase 0 audit plus the six-phase path

  1. Inventory and drift auditInspect existing skills, roles, profiles, adapters, conflicts, and stale artifacts before mutation.
  2. Domain analysisInspect the project, task types, constraints, and existing guidance.
  3. Team architecture designSelect the smallest coordination pattern that preserves quality.
  4. Role and artifact definitionName responsibilities, write ownership, inputs, outputs, and failure policy.
  5. Skill generationWrite portable skills with clear selection boundaries and progressive disclosure.
  6. Integration and orchestrationConnect phases through concise summaries or durable `_workspace/` handoffs.
  7. Validation and testingRun normal and failure scenarios, then simplify anything that does not earn its weight.

Read the detailed Phase 0 audit and six-phase workflow guide for phase outputs and stop conditions.

Choose the coordination shape

Pattern Best when Start here
Pipeline each phase depends on the previous artifact Pipeline
Fan-out/Fan-in independent read-heavy work benefits from parallel coverage Fan-out/Fan-in
Expert Pool a router should select a specialist for each request Expert Pool
Producer-Reviewer an output needs an explicit critique and bounded revision Producer-Reviewer
Supervisor a changing backlog needs central prioritization and reassignment Supervisor
Hierarchical Delegation a large goal naturally separates into shallow sub-goals Hierarchical Delegation

[!TIP] Keep tightly coupled work in one context. Add delegation only when the boundary, ownership, and synthesis responsibility can be stated clearly.

Keep the portable source visible

The public portal explains the workflow. The repository remains the authority for the reusable skill and its contracts:

Surface Purpose
.agents/skills/harness/SKILL.md canonical Phase 0 audit, six-phase workflow, and portable defaults
docs/harness/ durable team-spec and role-artifact contracts
_workspace/ deterministic intermediate handoffs when inspection or resumption matters
scripts/ installer and repository validation

For implementation details, read the canonical Harness skill on GitHub.

Continue

For maintainers

Review the artifact contract, starter example, and repository validation commands.

Maintainer reference →