author-workflow

Multi-Agent Harness

Generic multi-agent harness for turning a project or business domain into an agent-operable system.

The harness starts from a goal, models the domain scenario workflow, identifies the missing infrastructure that would shorten agent work, designs the right agent team and task graph, and then drives execution through messages, evidence-backed reports, critic review, decisions, and follow-up requirements.

This repository is intentionally separate from any specific trading, research, or product codebase. A project such as Earning Engine should be integrated through an adapter: skills and tool descriptors teach agent members how to use that project's CLI, Dashboard, artifacts, and permission rules.

Product Boundary

Multi-Agent Harness
  Goal / AgentTeam / AgentMember / AgentRuntime / AgentEvent / Task / Message
  Proposal / Evidence / Decision / ProviderSession
  Skill files / Tool descriptors / Agent Dashboard

Project Adapter
  CLI commands / Dashboard links / artifact readers / domain acceptance /
  permissions / evidence policy

The generic core must not import project-specific runtime code.

Future objects such as Report, Claim, Blocker, and Permission are not first-version gateable contracts until they have schemas, implementation, and checks.

Quickstart

Install the author-workflow skill into your agent, start the harness service, then ask your agent to author and run a workflow. Full walkthrough: docs/getting-started.md.

# 1. install the skill (Claude Code + Codex)
scripts/install-skill.sh --agent both
#    or standalone:  curl -fsSL .../scripts/install-skill.sh | bash -s -- --agent both
#    or:             npx skills add cyl19970726/multi-agent-harness --skill author-workflow --agent codex
#    or (Claude):    /plugin marketplace add cyl19970726/multi-agent-harness && /plugin install author-workflow

# 2. start the service
cargo build -p harness-cli
./target/debug/harness serve --addr 127.0.0.1:8787   # API + store
pnpm install && pnpm dashboard:dev                    # dashboard UI (watch runs live)

# 3. run a workflow your agent authored
./target/debug/harness workflow run-script prog.star \
    [--timeout-ms 300000] [--max-budget-usd 2.00] [--resume <prior_run_id>]

Verify the whole install → service → run journey at any time:

pnpm acceptance:skill-install            # local checks (no network)
pnpm acceptance:skill-install --remote   # + the anonymous curl|bash install path

One serve / dashboard can manage many projects (each with its own goals/tasks/runs in a centralized store under ~/.harness/projects/<id>/) plus a reserved GLOBAL ~/ project. Register with harness init, switch with harness project switch <id|path>, migrate a legacy repo-local .harness with harness project migrate. See docs/multi-project.md.

Repository Layout

Path	Purpose
docs/	PRD, architecture, operations, schemas, and decisions.
schemas/	Stable JSON schemas shared by API, CLI, adapters, and Dashboard.
crates/	Rust backend crates.
apps/agent-dashboard	React/Vite Agent Dashboard control-plane app and static build output.
skills/	Shipped, installable skills (the author-workflow deliverable).
.agents/skills/	This repo's internal runtime skills (auto-discovered by Codex / harness-spawned workers).
examples/adapters/earning-engine	First project adapter example.

Start Here

• Agent operating rules
• Product requirements
• MVP
• Design basis
• Concept Model
• Architecture
• Core Modules
• Data Model
• Agent Runtime
• Agent Control Plane
• Agent Dashboard
• Agent Dashboard Frontend Architecture
• Agent Dashboard Runbook
• Git / PR Workflow
• Multi-Project Harness
• Provider Integrations
• Goal Learning Loop
• Codex Integration
• Operations
• Schemas
• Decisions

Skills

Shipped (install into your own project — see the Quickstart):

• Author workflow: teach a shell-capable agent (Claude Code / Codex) to author a Starlark multi-agent workflow and run it with harness workflow run-script. Install with scripts/install-skill.sh --agent both.

Internal (this repo's own agents use these via .agents/skills/):