okf-agent-graph

OKF Agent Graph Engineering Runtime (AGER) — a Claude Code, Codex, Grok Build, Grok Bot, and LangChain Deep Agents plugin for authoring multi-agent loop graphs as OKF (Markdown + YAML).

Multi-host

Write isolation: docs/ISOLATION.md. LoopPolicy / KnowledgeBind read main and write brain/<actor>/<session-id>. Public examples use fictional lumenfield-detector and northstar-console only.

Also: Onboarding · skill ager-session.

Why this plugin

okf-graph-eng gives you portable OKF graphs, impact analysis, validation, and progressive disclosure.

okf-agent-graph adds the domain model for multi-agent runtime config:

• Orchestrator / Doer / Judge / Synthesizer roles
• LoopControls: goal, deadline, price, max_turns, no_progress
• Tools with expression block rules, idempotency, secrets
• ScratchPad KV (set / append lists) + lineage
• Run + Trigger, FailurePolicy, RetryPolicy, Compensation
• KnowledgeBind / RetrievalBinding (OKF / project knowledge)

Use as a spec and as runtime-loadable config (adapters map to LangGraph, CrewAI, OpenAI Agents, Anthropic patterns).

Nouns (this plugin)

AGER owns agent / harness / loop / runtime / eval types. Catalog and ContextPack live in okf-plugin.

AgentNode, OrchestratorAgent, WorkerAgent, JudgeAgent, SynthesizerAgent, RouterAgent, GuardrailAgent, HumanGate, AgentGraph, AgentGraphModule, Workflow, FanOut, FanIn, ParallelGroup, ControlEdge, Harness, SharedState, ToolCapability, Tool, ToolRule, LoopPolicy, LoopControl, ScratchPad, LineageRecord, EpisodeStore, KnowledgeBind, RetrievalBinding, SharedChannel, MemoryArtifact, ContextIsolationPolicy, InputSchema, OutputSchema, Run, Trigger, FailurePolicy, RetryPolicy, Compensation, CircuitBreaker, DeadLetter, CheckpointPolicy, RunTrace, StreamPolicy, HandoffPolicy, Rubric, Criterion, Judgment, EvalSuite, SecretRef, RateLimit, ConcurrencyLimit, DataClassPolicy.

RateLimit is also a SAC config/gateway noun. AGER RateLimit is runtime quota on a tool or loop.

Reverse engineering (AGKC): point the plugin at an existing agent repo — Claude/ChatGPT API, LangChain, LangGraph, CrewAI, LlamaIndex, Claude Agent SDK, Deep Agents, MCP, Bedrock AgentCore, hardened microVMs/containers — and extract prompts, tools/JSON-RPC/MCP schemas, orchestration graphs, loop/harness policies, and runtime isolation into draft AGER knowledge. Same shape as system-architecture-capture and data-engineering-knowledge-capture, for the agent-graph domain.

Dependency (required)

Install okf-plugin first — this plugin reuses its graph validate/impact/query tooling.

# 1) Core OKF graph engineering
claude plugin marketplace add SpillwaveSolutions/okf-plugin
claude plugin install okf-graph-eng@okf-plugin-marketplace

# 2) Agent graph / AGER domain
claude plugin marketplace add SpillwaveSolutions/okf-agent-graph
claude plugin install okf-agent-graph@okf-agent-graph-marketplace

For Codex, add the same repository marketplace, open /plugins, install okf-agent-graph, and start a new session:

codex plugin marketplace add SpillwaveSolutions/okf-agent-graph
codex
# /plugins

The pinned okf-plugin v0.3.2 dependency does not ship a Codex manifest. Run full OKF validation through an okf executable, a sibling ../okf-plugin checkout, or an explicitly supplied path to okf-graph.py.

Grok Build loads Claude-compatible plugins with zero extra config.

Dependency contract

See docs/DEPENDENCIES.md.

Quick start

1. Ensure okf-graph-eng is installed
2. /ager-init or $ager-init-graph — scaffold an AGER bundle
3. /ager-author or $ager-author — add Orchestrator, Workers, Judge, LoopPolicy, Tools
4. /ager-validate or $ager-validate — AGER rules + delegate to okf validate
5. /ager-pack — tiny or 2-hop ContextPack. Fail-closed over the token budget.
6. Optional: /ager-compile — emit adapter notes / stubs for a target framework

The scripts are also directly callable:

python3 scripts/ager-init.py agent-graph --title "My agent graph"
python3 scripts/ager-validate.py agent-graph --strict

# Reverse engineer an existing agent project
python3 scripts/ager_scan.py --root /path/to/agent-project --json -o scan.json
python3 scripts/ager_reverse_engineer.py --root /path/to/agent-project --out discovered-ager --json

Sample self-describing graph:

# Requires okf-plugin scripts on path or clone sibling
python3 scripts/ager-validate.py sample-ager --strict
python3 ../okf-plugin/scripts/okf-graph.py validate sample-ager --strict

Skills & commands

Codex invokes the same skills with $ager-init-graph, $ager-author, $ager-validate, $ager-compile, $ager-scan, and $ager-reverse-engineer.

Agent

• agent-graph-engineer — designs multi-agent loops using AGER + okf impact/pack
• agent-graph-re-orchestrator — reverse-engineers existing agent codebases into AGER (AGKC)

Concept planes (AGER v0.3)

1. Core — AgentNode roles, AgentGraph, FanOut/In, modules, I/O schemas
2. Control — LoopPolicy / LoopControl (goal, deadline, price, max_turns, no_progress)
3. Memory — ScratchPad KV + lineage, EpisodeStore, KnowledgeBind, RetrievalBinding
4. Ops / action — Run, Trigger, Failure/Retry/Compensation, Tool+ToolRule, SecretRef, quotas

Full catalog: docs/AGER_SPEC.md · sample: sample-ager/

Ecosystem map

okf-plugin (okf-graph-eng)     portable OKF graph ops
        ▲
        │ depends on
okf-agent-graph (this)         multi-agent loop domain
        │
        ├── wiki_ticket_sdd    Trigger kind: ticket_event
        └── knowledge/ (PKC)   KnowledgeBind source

License

MIT — see LICENSE.