AWP — Agent Workspace Protocol

The portable, open format for AI agent identity, memory, and behavior.

AWP defines what an agent is — its persistent state, behavioral constraints, accumulated knowledge, reputation, and coordinated work — in framework-agnostic, version-controlled, human-readable files.

your-agent/
  .awp/workspace.json    # Workspace manifest
  IDENTITY.md            # Who the agent is (name, capabilities, DID)
  SOUL.md                # How the agent behaves (values, boundaries, governance)
  USER.md                # Human profile (optional)
  AGENTS.md              # Operational instructions (optional)
  TOOLS.md               # Environment config (optional)
  memory/                # Structured daily logs
    2026-01-30.md
  artifacts/             # Versioned knowledge artifacts (SMP)
    llm-research.md
  reputation/            # Multi-dimensional reputation profiles (RDP)
    research-bot.md
  contracts/             # Delegation contracts (RDP)
    q3-research.md
  projects/              # Coordination projects and tasks (CDP)
    q3-launch.md
    q3-launch/tasks/
      competitive-analysis.md
  swarms/                # Multi-agent composition (CDP)
    q3-launch-team.md

Every file is valid Markdown (human-readable) with structured YAML frontmatter (machine-parseable). No database, no server, no runtime. Just files in a directory, compatible with Git by default.

Quick Start

npm install -g @agent-workspace/cli

# Scaffold a new agent workspace
awp init

# Check it's valid
awp validate

# See what you've got
awp inspect

# Generate a DID identity
awp identity generate

# Export as A2A Agent Card
awp identity export

Why AWP?

A2A owns transport. MCP owns tools. ACK owns payments. Nobody owns the definition of the agent itself.

AWP sits below all other protocols — defining what the agent is, while they define what the agent does.

Packages

Dashboard

Visual governance layer for AWP workspaces. Reads workspace files from disk and displays agent identity, projects, tasks, reputation, artifacts, contracts, and memory.

# Run the dashboard (point to a workspace)
AWP_WORKSPACE=/path/to/workspace npm run dev --workspace=packages/awp-dashboard

# Opens at http://localhost:3000

Pages:

• Overview — Agent identity, health warnings, workspace metrics, active tasks
• Projects — Project list with progress bars, detail pages with kanban task boards
• Reputation — Agent roster with score gauges, profile pages with radar charts
• Artifacts — Knowledge browser with confidence bars, detail with provenance timelines
• Contracts — Delegation contracts with evaluation scores
• Memory — Daily log timeline with long-term memory

MCP Server

Plug any AWP workspace into Claude Code, Cursor, or any MCP client:

# Add to Claude Code
claude mcp add awp-workspace -- npx @agent-workspace/mcp-server

# Or with a specific workspace path
AWP_WORKSPACE=/path/to/workspace npx @agent-workspace/mcp-server

Tools exposed:

• awp_read_identity — Read agent identity
• awp_read_soul — Read behavioral constraints
• awp_read_user — Read human profile
• awp_read_memory — Read memory (daily or long-term)
• awp_write_memory — Log structured memory entries
• awp_artifact_read — Read a knowledge artifact
• awp_artifact_write — Create or update a knowledge artifact
• awp_artifact_list — List all artifacts
• awp_artifact_search — Search artifacts
• awp_reputation_query — Query agent reputation (with time-based decay)
• awp_reputation_signal — Log a reputation signal
• awp_contract_create — Create a delegation contract
• awp_contract_evaluate — Evaluate a contract and generate reputation signals
• awp_project_create — Create a coordination project
• awp_project_list — List projects with status filter
• awp_project_status — Project details with task summary
• awp_task_create — Create a task within a project
• awp_task_update — Update task status or assignee
• awp_task_list — List tasks for a project
• awp_task_graph — Analyze task dependencies (topological sort, cycles, critical path)
• awp_artifact_merge — Merge artifacts (additive or authority strategy)
• awp_swarm_create — Create a multi-agent swarm
• awp_swarm_list — List swarms with status filter
• awp_swarm_show — Show swarm details and staffing
• awp_swarm_recruit — Find candidates for unfilled roles
• awp_swarm_assign — Assign agent to swarm role
• awp_swarm_role_add — Add a role to a swarm
• awp_workspace_status — Workspace summary with health warnings

Reputation & Delegation (RDP)

Track agent trustworthiness and define work agreements:

# Log a reputation signal for an agent
awp reputation signal research-bot \
  --agent-did did:key:z123 --agent-name ResearchBot \
  --dimension reliability --score 0.9 --message "Delivered on time"

# Query reputation (scores decay over time toward 0.5)
awp reputation query research-bot

# List all tracked agents
awp reputation list

# Create a delegation contract
awp contract create q3-research \
  --delegate did:key:z123 --delegate-slug research-bot \
  --description "Research LLM context window techniques"

# Evaluate a completed contract (generates reputation signals automatically)
awp contract evaluate q3-research \
  --completeness 0.9 --accuracy 0.85 --clarity 0.8 --timeliness 1.0

Reputation uses EWMA (exponentially weighted moving average) with time-based decay — scores degrade without new signals, confidence tracks sample size independently, and domain competence is a separate map. See spec/rdp/rdp-spec.md for the full specification.

Coordination (CDP)

Organize multi-agent work with projects, tasks, and reputation-gated assignments:

# Create a project
awp project create q3-launch --title "Q3 Product Launch" --deadline 2026-04-01

# Add tasks with assignees and reputation gates
awp task create q3-launch research \
  --title "Competitive Analysis" \
  --assignee did:key:z123 --assignee-slug research-bot \
  --priority high --deadline 2026-02-15

# Create a blocked task
awp task create q3-launch positioning \
  --title "Positioning Doc" \
  --blocked-by task:q3-launch/research

# Update task status
awp task update q3-launch research --status in-progress

# Rich workspace overview
awp status

# Authority-based artifact merge (uses reputation scores)
awp artifact merge target-slug source-slug --strategy authority

Project members can have minReputation thresholds — when assigning a task, the system checks the assignee's reputation and warns if below threshold.

Task Dependencies

Analyze task dependency graphs with topological sorting, cycle detection, and critical path analysis:

# Analyze dependencies for a project
awp task graph q3-launch

# Check for dependency cycles (exits with error if found)
awp task graph q3-launch --check

# JSON output for programmatic use
awp task graph q3-launch --json

Swarms (Multi-Agent Coordination)

Compose multi-agent teams with reputation-gated role recruitment:

# Create a swarm
awp swarm create q3-launch-team \
  --name "Q3 Launch Team" \
  --goal "Complete Q3 product launch" \
  --project q3-launch

# Add roles with reputation requirements
awp swarm role add q3-launch-team researcher \
  --count 1 \
  --min-reputation reliability:0.7 \
  --min-reputation domain-competence:ai-research:0.8

awp swarm role add q3-launch-team writer \
  --count 2 \
  --min-reputation technical-writing:0.7

# Find qualified candidates from reputation profiles
awp swarm recruit q3-launch-team

# Auto-assign best qualified candidates
awp swarm recruit q3-launch-team --auto

# View swarm status
awp swarm show q3-launch-team

# List all swarms
awp swarm list
awp swarm list --status recruiting

Swarms automatically transition from recruiting to active status when all roles are filled. See spec/cdp/cdp-spec.md for the Coordination Protocol specification.

Experiments (EXP)

Run manifesto-driven experiments with multi-agent societies:

# Create a society of 3 agents
awp experiment society create --manifesto MANIFESTO.md --agents 3 --seed 42

# Run 5 cycles
awp experiment run --society manifesto-xxx-123 --cycles 5 --manifesto MANIFESTO.md

# Use Anthropic instead of OpenAI
awp experiment run --society manifesto-xxx-123 --cycles 5 --manifesto MANIFESTO.md --provider anthropic

# List all societies
awp experiment list

# Show society details and results
awp experiment show manifesto-xxx-123

Environment Variables

# For OpenAI agents (default)
export OPENAI_API_KEY=sk-xxx

# For Anthropic agents
export ANTHROPIC_API_KEY=sk-ant-xxx

Moltbot / Clawdbot Integration

AWP experiments can be run through moltbot (clawdbot) for conversational access:

# Install the skill
cp -r skills/awp-experiment ~/.nvm/versions/node/*/lib/node_modules/clawdbot/skills/

# Link the CLI globally
npm link -w packages/awp-cli

# Use via clawdbot
clawdbot agent --message "Create a 3-agent AWP society"
clawdbot agent --message "Run 2 cycles on society manifesto-xxx"

Programmatic Usage

import { 
  OpenAIAgent, 
  AnthropicAgent,
  ExperimentOrchestrator, 
  SocietyManager,
  parseManifesto 
} from '@agent-workspace/agent';

// Parse manifesto
const manifesto = await parseManifesto('./MANIFESTO.md');

// Create society
const manager = new SocietyManager('./societies');
const society = await manager.createSociety('my-experiment', manifesto.id, 3);

// Create agents (OpenAI or Anthropic)
const agents = society.agents.map((ws, i) => 
  new OpenAIAgent(`agent-${i}`, ws, 'gpt-4o-mini')
  // Or: new AnthropicAgent(`agent-${i}`, ws, 'claude-sonnet-4-20250514')
);

// Run experiment
const orchestrator = new ExperimentOrchestrator(manifesto, agents, new MetricsCollector(), society);
const results = await orchestrator.runExperiment(10);

console.log(`Success rate: ${results.aggregateMetrics.overallSuccessRate * 100}%`);

See packages/awp-agent/AGENT_ADAPTER.md for implementing custom agent adapters.

File Format

AWP files are dual-format — valid Markdown with YAML frontmatter:

---
awp: "0.4.0"
type: "soul"
vibe: "Competent, dry wit, gets stuff done"
values:
  - id: "genuine-helpfulness"
    priority: 1
    description: "Be genuinely helpful, not performatively helpful"
boundaries:
  - id: "privacy"
    rule: "Private things stay private. Period."
    severity: "hard"
governance:
  humanApprovalRequired:
    - "sending emails"
    - "posting publicly"
  autonomouslyAllowed:
    - "reading files"
    - "updating memory"
---

# SOUL.md - Who You Are

Be genuinely helpful, not performatively helpful. Skip the "Great question!"
and "I'd be happy to help!" — just help.

Identity & Interop

AWP uses W3C DIDs for agent identity:

# Generate a did:key identity
awp identity generate

# Export as A2A-compatible Agent Card (JSON)
awp identity export

The exported Agent Card is compatible with Google's Agent-to-Agent protocol, so AWP agents can participate in A2A networks.

Memory

Structured daily logs with machine-parseable entries:

# Log a memory entry
awp memory log --tag learning --tag architecture \
  "Discovered that the auth service uses JWT with RS256"

# Search memory
awp memory search "auth"

Memory entries support pinned: true to prevent compaction of critical knowledge.

Knowledge Artifacts (SMP)

Versioned, provenanced knowledge documents with confidence scores:

# Create an artifact
awp artifact create llm-research --title "LLM Context Research" --tags ai,research

# Edit the markdown body, then commit the change
awp artifact commit llm-research -m "Added benchmark comparison"

# List all artifacts
awp artifact list

# Search by content, title, or tags
awp artifact search "context window"

# View provenance history
awp artifact log llm-research

# Merge one artifact into another
awp artifact merge target-slug source-slug
awp artifact merge target-slug source-slug --strategy authority

Artifacts track who wrote what, when, and with what confidence — full provenance for agent knowledge. See spec/smp/smp-spec.md for the Shared Memory Protocol specification.

Specification

The full protocol specification is at spec/awp-spec.md. Sub-protocol specs: SMP (Shared Memory), RDP (Reputation & Delegation), CDP (Coordination). JSON schemas for all file types are in spec/schemas/.

Design Principles

1. Dual-format — Human-readable Markdown + machine-parseable YAML frontmatter
2. File-native — A workspace is a directory. Git-compatible by default
3. Framework-agnostic — Works with Claude Code, CrewAI, LangGraph, AutoGen, or custom agents
4. Incrementally adoptable — Start with just IDENTITY.md, add files as needed
5. Interoperable — DIDs, A2A Agent Cards, MCP tools

License

Apache-2.0