Efficient Agent Protocol (EAP)

Status: v1.0 Release Candidate. Core APIs and schema are frozen per docs/v1_contract.md. See STABILITY.md and ROADMAP.md for guarantees and planned milestones. Latest stable release: v1.0.1

Efficient Agent Protocol is a local-first framework for multi-step tool workflows. It stores large outputs as pointer-backed state (ptr_*) and runs dependency-aware DAG steps in parallel. It also ships OpenClaw interop paths for gateway/tool integration.

30-Second Pitch

• EAP is the reliability layer for agent workflows: deterministic execution, resumable runs, and pointer-backed state.
• It is built for local-first teams that need control over failure behavior, replayability, and traceability.
• It integrates with existing ecosystems instead of forcing a rewrite (chat_completions, responses, OpenClaw tooling, MCP tools).

See docs/eap_proof_sheet.md for reproducible evidence and command-level validation. Current patch-release evidence is tracked in docs/releases/v1.0.1-readiness.md.

What We Closed Recently

• OpenClaw interoperability moved from spike to shipped paths:
  • plugin adapter + skill pack
  • gateway /tools/invoke bridge
  • agent-routing header support
• Runtime control plane shipped with scoped auth and ownership governance:
  • execute/run-summary/pointer-summary/resume endpoints
  • policy guardrails with deterministic JSON errors
• Reliability hardening is now CI-enforced:
  • eval scorecard gate
  • competitive benchmark lane
  • soak + chaos reliability gate with regression thresholds
• Crash-safe resume/replay and MCP tool bridge are implemented and covered by integration tests.
• Self-hosted reference stack, telemetry export pack, and runbooks are in-repo and exercised in CI smoke lanes.

Demo Flow

Short flow: chat request -> pointer inspection -> trace/HITL checkpoint.

Architecture At A Glance

High-level runtime architecture (planning -> execution -> pointers/state -> API/UI).

Why Choose EAP

Where EAP Fits

Best for:

• Python developers building local-first orchestration with explicit execution semantics.
• Teams that care about observability, replayability, and controlled failure behavior.

Not ideal for:

• Teams expecting a fully managed hosted control plane (instead of self-hosting runtime components).
• Non-technical users wanting no-ops onboarding without any runtime/provider configuration.

Current Limits (Honest)

• Core APIs and schema are frozen for v1 (see docs/v1_contract.md). Surfaces marked unstable may still change.
• responses streaming behavior depends on gateway SSE support; see docs/streaming_compatibility.md for the 9-gateway compatibility matrix and known quirks.
• Performance/reliability thresholds are calibrated from repo baselines; production teams should tune them for their own workloads.
• This remains an engineering-first runtime, not a no-code orchestration product.

What You Get

• Pointer-based state to keep prompts small
• Parallel DAG execution with retries and validation
• Human-in-the-loop checkpoints (approval_required, approved, rejected)
• Crash-safe resume/replay from persisted run checkpoints
• Evaluation harness with CI threshold gates (scripts/eval_scorecard.py)
• Operator telemetry pack export (scripts/export_telemetry_pack.py)
• Self-hosted control-plane reference stack (deploy/self_hosted/docker-compose.yml)
• Scoped runtime auth + run ownership governance for remote operations
• Built-in chat UI (Streamlit) with trace + data inspection
• Conversation memory (full/window/summary)
• Fluent workflow builder API (WorkflowBuilder) for declarative step definitions
• Guided onboarding wizard (scripts/eap_onboarding_wizard.py) for new users
• Pluggable pointer storage backends (SQLite, Redis, PostgreSQL)
• OpenClaw and MCP interop:
  • OpenAI-compatible modes: chat_completions and responses
  • Gateway tool bridge for POST /tools/invoke
  • OpenClaw plugin + skills starter package in integrations/openclaw/eap-runtime-plugin
  • MCP tool bridge (invoke_mcp_tool)

Why Not Just Use A Generic Agent Framework?

• If you mainly need prompting convenience and managed UX, a platform suite may be simpler.
• If you need explicit run-state contracts, replayability, and pointer-backed payload discipline, EAP is a stronger fit.
• If you need to integrate with OpenClaw without rewriting your runtime, EAP now has first-party bridge paths.

Quickstart (GitHub-first)

Requirements:

• Python 3.10-3.13 (3.11 recommended)

Recommended one-command bootstrap (macOS/Linux):

git clone https://github.com/GenieWeenie/efficient-agent-protocol.git
cd efficient-agent-protocol
./scripts/bootstrap_local.sh

Expected output includes:

• Smoke workflow succeeded.
• Trace artifact: .../artifacts/bootstrap/bootstrap_trace.json

Windows fallback:

• Use WSL2 (Ubuntu) and run the same bootstrap command inside WSL.
• If you are not using WSL2, follow the manual setup path below.

Guided onboarding wizard (interactive):

pip install -e .
python scripts/eap_onboarding_wizard.py

Manual setup (cross-platform):

1. Install package

pip install -e .

2. Configure

python scripts/eap_doctor.py init-env --output .env --force

Minimum variables:

EAP_BASE_URL=http://localhost:1234
EAP_MODEL=nemotron-orchestrator-8b
EAP_API_KEY=not-needed

3. Smoke test

python -m examples.01_minimal

4. Run doctor diagnostics

python scripts/eap_doctor.py doctor --env-file .env --output-json artifacts/doctor/diagnostics.json

5. Run dashboard

pip install streamlit pandas
streamlit run app.py  # from the repository root

6. Use it

• Open http://localhost:8501
• In Agent Chat, ask for a task
• Check Data Inspector for pointer payloads
• Check Execution Trace for step timing/retries/errors

7. Try starter packs

python -m starter_packs.research_assistant --question "What are launch risks?"
python -m starter_packs.doc_ops --focus "summarize priorities and actions"
python -m starter_packs.local_etl

8. Optional OpenClaw smoke check

./scripts/interop_openclaw_smoke.sh v2026.2.22

9. Optional self-hosted reference stack

cp deploy/self_hosted/.env.example deploy/self_hosted/.env
docker compose --env-file deploy/self_hosted/.env -f deploy/self_hosted/docker-compose.yml up --build -d
python scripts/self_hosted_stack_smoke.py --base-url http://127.0.0.1:8080 --bearer-token "<runtime-token>"

10. Architecture + extension deep dives

• docs/architecture.md
• docs/custom_tool_authoring.md
• docs/pointer_internals.md

Programmatic Example

import asyncio
from eap.protocol import StateManager
from eap.environment import AsyncLocalExecutor, ToolRegistry
from eap.environment.tools import read_local_file, READ_FILE_SCHEMA
from eap.agent import AgentClient

state_manager = StateManager()
registry = ToolRegistry()
registry.register("read_local_file", read_local_file, READ_FILE_SCHEMA)
executor = AsyncLocalExecutor(state_manager, registry)

architect = AgentClient(
    base_url="http://localhost:1234",
    model_name="nemotron-orchestrator-8b",
    provider_name="local",
)

manifest = registry.get_agent_manifest()
macro = architect.generate_macro("Read README.md and summarize setup steps", manifest)
result = asyncio.run(executor.execute_macro(macro))
print(result)

Common Commands

./scripts/bootstrap_local.sh
python scripts/eap_doctor.py init-env --output .env --force
python scripts/eap_doctor.py doctor --env-file .env --output-json artifacts/doctor/diagnostics.json
python3 -m pytest -q
pre-commit run --all-files
python3 scripts/migrate_state_db.py --db-path agent_state.db --dry-run
python3 scripts/export_metrics.py --db-path agent_state.db --output metrics/latest.json
python3 scripts/export_telemetry_pack.py --db-path agent_state.db --output-dir artifacts/telemetry
python scripts/eap_state_backup.py backup --db-path agent_state.db --output-root artifacts/state_backups
python scripts/soak_chaos_scorecard.py --output-dir artifacts/soak_chaos --threshold-config docs/soak_chaos_thresholds.json --baseline docs/soak_chaos_baseline.json
./scripts/interop_openclaw_smoke.sh v2026.2.22
docker compose --env-file deploy/self_hosted/.env -f deploy/self_hosted/docker-compose.yml up --build -d
python scripts/self_hosted_stack_smoke.py --base-url http://127.0.0.1:8080 --bearer-token "<runtime-token>"
python3 -m build

Docs

Full documentation index: docs/README.md

• Start here:
  • docs/eap_proof_sheet.md
  • docs/configuration.md
  • docs/architecture.md
  • docs/custom_tool_authoring.md
  • docs/pointer_internals.md
  • docs/troubleshooting.md
• Contract and policy:
  • STABILITY.md
  • ROADMAP.md
  • docs/v1_contract.md
  • docs/typing_policy.md
  • SECURITY.md
  • CONTRIBUTING.md
• Runtime and operations:
  • docs/workflow_schema.md
  • docs/tools.md
  • docs/observability.md
  • docs/operator_telemetry_pack.md
  • docs/state_backup_restore.md
  • docs/soak_chaos_reliability.md
  • docs/self_hosted_control_plane.md
  • docs/remote_ops_governance.md
  • docs/migrations.md
• Release evidence:
  • docs/releases/v1.0.1-readiness.md
  • docs/release.md
  • docs/github_actions_runtime_inventory.md
• Interop and starter packs:
  • docs/openclaw_interop.md
  • integrations/openclaw/eap-runtime-plugin/README.md
  • docs/starter_packs/README.md
• GitHub roadmap board: https://github.com/users/GenieWeenie/projects/1