AEF - Agent Evaluation Framework

AEF is a framework to generate tests, run/evaluate trajectories, collect feedback, and self-evolve agent behavior.

The workflow is intentionally minimal and framework-agnostic:

• aef generate calls the generation component/tool
• aef evaluate calls the evaluation component/tool
• aef feedback calls the feedback component/tool
• aef evolve calls the evolution component/tool

Internally, these are routed through an A2A bus so the same flow works for sub-agents implemented with different frameworks.

---

Installation

From PyPI

Install via pip or uv:

pip install aef-framework

or with uv:

uv pip install aef-framework

Local Development Install with uv

AEF uses uv for fast, reliable Python package management.

1. Install uv (if not already installed)

curl -LsSf https://astral.sh/uv/install.sh | sh

2. Create a virtual environment

cd AEF
uv venv --python=3.11

This creates a .venv directory with Python 3.11 (or use 3.10, 3.12 as needed).

3. Activate the virtual environment

source .venv/bin/activate  # Linux/macOS
# or
.venv\Scripts\activate     # Windows

4. Install AEF in editable mode

uv pip install -e .

This installs AEF and all dependencies, making the aef command available.

5. Verify installation

aef --help

Traditional pip install (local)

If you prefer using pip:

python -m venv .venv
source .venv/bin/activate
pip install -e .

---

Core Principles

• Universal sub-agent support via adapter contract (python, cli, http)
• Single essential loop: Generate → Evaluate → Feedback → Evolve
• Composable A2A components instead of tightly-coupled command logic
• Versioned evolution profiles with before/after evaluation comparison

---

Basic Workflow

1) Generate trajectories

aef generate --config configs/fleet_ccc_run.json --n 10

2) Evaluate against a golden run

aef evaluate --config configs/fleet_ccc_run.json --golden run_YYYYMMDD_xxxxxx

3) Submit feedback

aef feedback --agent fleet_ccc --text "Agent should ask confirmation before delete operations"

4) Evolve (auto-apply + compare)

aef evolve --config configs/fleet_ccc_run.json --n 10

aef evolve now performs:

1. baseline evaluate
2. classify feedback into amendments
3. apply evolution profile
4. re-evaluate and report before/after score delta

---

Use AEF With Any Sub-Agent

Set agent.adapter_type in your config:

• python: ADK/Python agent entrypoint module_or_file.py:agent_var
• cli: shell command template using {step} / {goal} placeholders
• http: endpoint that accepts { goal, step, session_id? }

See detailed usage in docs/USING_ANY_SUBAGENT.md.

Runtime endpoint mode (--agent_endpoint)

In addition to config-defined adapters, you can override execution at runtime:

• If --agent_endpoint is provided, AEF routes AUT calls through HTTP endpoint mode.
• If --agent_endpoint is not provided, AEF keeps existing behavior (for example local --sub_agent / config adapter).

Endpoint-mode guarantee:

• AEF uses the hosted endpoint runner path for AUT execution.
• Local Python entrypoint loading is not required in this mode.
• Local entrypoint import/path issues do not block endpoint-mode execution.
• Endpoint runner uses ADK server contract: create/reuse session, then call POST /run.

This is useful when the same agent can run locally in development and remotely in a hosted ADK/A2A service.

Examples:

# Generate through hosted endpoint
aef generate --config configs/fleet_ilo_run.json \
	--agent_endpoint http://localhost:8086/docs/ --n 2

# Evaluate through hosted endpoint
aef evaluate --config configs/fleet_ilo_run.json \
	--agent_endpoint http://localhost:8086/docs/ --golden run_YYYYMMDD_xxxxxx

# Evolve through hosted endpoint
aef evolve --config configs/fleet_ilo_run.json \
	--agent_endpoint http://localhost:8086/docs/ --n 5

Notes:

• /docs/ URLs are supported (AEF resolves them to API base).
• Endpoint mode is available on generate, evaluate, run, and evolve.
• Endpoint mode is intended for ADK/A2A-hosted AUTs where the AUT is reachable over API.

ADK endpoint contract used by --agent_endpoint:

• Session bootstrap: POST /apps/{app_name}/users/{user_id}/sessions/{session_id}
• Inference: POST /run
• Request payload: {"appName","userId","sessionId","newMessage":{"role":"user","parts":[{"text":...}]},"streaming":false}
• A 409 Conflict during session creation is treated as expected session reuse.

Trajectory logging in endpoint mode:

• steps[].content stores assistant text response.
• steps[].tool_calls, steps[].tool_responses, and steps[].tools_used store tool trace data when present.

Full prerequisites and onboarding checklist:

• docs/ADOPTING_NEW_AGENT.md

---

A2A Components

AEF components exposed through the internal bus:

• generation.generate
• evaluation.evaluate
• feedback.submit_text
• feedback.submit_annotations
• evolution.evolve

See docs/A2A_COMPONENTS.md.

---

Evolution Outputs

Evolution applies and versions runtime amendments per agent under:

• prompts/evolution_profiles/<agent>/latest.json
• prompts/evolution_profiles/<agent>/profile_<timestamp>.json

These profiles contain:

• prompt addenda
• tool policies
• generator hints
• agent hints
• rubric updates

See docs/SELF_EVOLUTION.md.

---

Web UI

AEF includes a Next.js web interface for managing agents, running benchmarks, reviewing trajectories, and tracking evolution.

Option A — Docker (recommended)

Run both the backend API and the web UI with a single command:

docker compose up -d

• Frontend: http://localhost:3010
• Backend API: http://localhost:8001

To rebuild after code changes:

docker compose up -d --build

To stop:

docker compose down

Your run database (aef_runs.db), configs, outputs, and annotated data are bind-mounted from the repo root and persist across restarts.

Option B — Local development

Start the backend:

uvicorn aef.api.main:app --reload --port 8001

Start the UI:

cd aef-ui
npm install
npm run dev

Open http://localhost:3010. See aef-ui/README.md for full details.

UI Pages

Page	Description
Dashboard	Score trends, model cost breakdown, runs-over-time — shows only COMPLETE runs
Agent Config	Register local Python agents or HTTP endpoints
Generate	Run trajectory generation with live progress, view past runs with expandable step-by-step conversations
Feedback	Review GENERATED trajectories with full multi-step detail, submit per-trajectory quality ratings
Evaluate	Select a golden trajectory run, re-execute and score against it with dimension radar charts
Evolve	Run self-improvement cycles on evaluated runs, manage memory deltas
Query	Browse runs, trajectories, evaluations and execute raw SQL

Pipeline Flow

Generate → (GENERATED) → Feedback → (COMPLETE) → Evaluate → (COMPLETE/REGRESSION) → Evolve

Each UI page filters to the appropriate pipeline stage so you only see runs relevant to that step.

---

Minimal Command Reference

# Generate
aef generate --config <config.json> --n 10

# Generate via hosted endpoint override
aef generate --config <config.json> --agent_endpoint <http://host:port/docs/> --n 10

# Direct A2A tool call
aef a2a --config <config.json> --component generation --tool generate --payload '{"n": 2}'

# Evaluate golden by run id
aef evaluate --config <config.json> --golden <run_id>

# Evaluate via hosted endpoint override
aef evaluate --config <config.json> --agent_endpoint <http://host:port/docs/> --golden <run_id>

# Feedback
aef feedback --agent <agent_name> --text "..."

# Evolve
aef evolve --config <config.json> --n 10

# Evolve via hosted endpoint override
aef evolve --config <config.json> --agent_endpoint <http://host:port/docs/> --n 10

# Compare two eval runs
aef compare --run <run_a> --vs <run_b>

# Query runs / memory
aef query runs --agent <agent_name>
aef query memory --agent <agent_name> --all-memory
aef query memory --agent <agent_name> --history

---

Documentation

• docs/AEF_WORKFLOW.md
• docs/A2A_COMPONENTS.md
• docs/USING_ANY_SUBAGENT.md
• docs/SELF_EVOLUTION.md
• docs/PUBLISHING.md - PyPI package publishing guide

---

Contributing

Contributions are welcome! See CONTRIBUTING.md for development setup and guidelines.

---

License

AEF is released under the Apache License 2.0. See LICENSE for details.