agentprop

AgentProp

Graph control for agent workflows.

AgentProp models AI-agent workflows as directed weighted graphs—agents, tools, context, verifiers, and failures become nodes and edges you can analyze, simulate, and supervise.

It is a control layer, not an orchestrator. Your harness emits one ExecutionEvent per step; AgentProp returns whether to continue, verify, switch strategy, or finalize. See the overview for the core ideas (metric-dimension verifiers, quality cascade, RZF scaling, runtime control).

Get started — pick your path

Path A: you use a coding agent (Claude Code / Codex)

Install the package and register the plugin or MCP server so your agent can analyze its own workflow and run under budget control:

python -m pip install "agentprop[mcp]"
agentprop doctor --tier graph

codex mcp add agentprop -- agentprop-mcp   # or the plugin bundle below

Then follow the beta quickstart. No API keys are needed for graph analysis.

Path B: you build multi-agent systems (LangGraph and friends)

python -m pip install agentprop

Analyze a LangGraph workflow in a few lines:

from agentprop import analyze

report = analyze(my_langgraph_workflow)
print(report.to_markdown())

Wrap it with runtime control:

from agentprop import wrap

controlled = wrap(my_langgraph_workflow, budget={"tokens": 100_000, "cost": 0.50})
result = controlled.run({"task": "ship the workflow"})
print(result.decision_trace)

Start with the tutorial and the framework status matrix (LangGraph, CrewAI, and OpenAI Agents have native builders; AutoGen and LlamaIndex are dict-interchange only).

Try it without any framework

agentprop analyze planner_coder_tester_reviewer
agentprop control-demo --demo terminal --out-dir reports/control-demo
agentprop view planner_coder_tester_reviewer --out reports/view.html

More: docs/index.md · troubleshooting · contributing

Development checkout:

python -m venv .venv && source .venv/bin/activate
python -m pip install -e ".[dev]"
make test

Early signal

On one Terminal-Bench 2.1 task (regex-log, Codex + gpt-5.5), AgentProp control preserved pass while reducing tokens and cost versus the raw agent:

That is a single-task early signal, not a benchmark claim. Multi-task replication is documented in Terminal-Bench multi-task protocol.

Coding-agent integration

AgentProp ships a same-repo plugin bundle at plugins/agentprop (Codex + Claude Code manifests, packaged skill, MCP config). Install the Python package, then register the plugin:

python -m pip install "agentprop[mcp]"

codex plugin marketplace add aryan5v/AgentProp --sparse .agents --sparse plugins
codex plugin add agentprop@agentprop

Portable skill-only install:

npx skills add aryan5v/AgentProp --skill agentprop-workflow-optimizer

Full setup, MCP registration, and troubleshooting: coding agents · plugin distribution.

Examples

python examples/coding_agent_full_suite.py --out-dir reports/full-suite
python examples/minimal_control_loop.py

More: examples/README.md.

Repository map

Details: repository layout.

Status

Public alpha research software. Graph analysis, propagation, runtime control, and key-free demos work without API keys. Treat live-agent numbers as directional until larger studies with saved artifacts are published under docs/results/.

License

Apache 2.0. See SECURITY.md and CHANGELOG.md.