开源AI工作流

Build subcommand (default)

poetry run python -m ai_team run "Build a minimal Flask API" \ --backend langgraph --team backend-api --env dev --skip-estimate

Quick start

```bash git clone https://github.com/RickZee/ai-team.git && cd ai-team cp .env.example .env # Add API keys (see Configuration below) poetry install

Demo projects

Six ready-to-run scenarios that exercise the full pipeline:

```bash poetry run python scripts/run_demo.py demos/01_hello_world poetry run python scripts/run_demo.py demos/02_todo_app --skip-estimate

Demo 06: AutoOptimizer Loop

ai-team optimize demos/06_karpathy_optimization/workspace \ --metric demos/06_karpathy_optimization/metric.yaml \ --strategy demos/06_karpathy_optimization/strategy.md \ --budget 2.00 ```

Each demo directory contains input.json with the project spec and expected_output.json as an acceptance contract. After a run, scripts/capture_demo.py verifies the output and writes RESULTS.md.

For the full file layout, schema reference, capture/verification workflow, and instructions for adding new demos, see docs/DEMOS.md.

Full options

ai-team optimize ./workspace/my-app \ --metric metric.yaml \ --strategy strategy.md \ --backend claude-agent-sdk \ --team research-optimizer \ --budget 5.00 \ --max-experiments 50 \ --branch optimize/my-run \ --editable src/ lib/

| Flag | Description | Default |
|------|-------------|---------|
| `--metric` | Path to `metric.yaml` (name, evaluation_command, direction) | required |
| `--strategy` | Path to a Markdown hints file for the optimizer agent | — |
| `--backend` | Backend to use | `claude-agent-sdk` |
| `--team` | Team profile | `research-optimizer` |
| `--budget` | Total USD budget across all experiments | `10.0` |
| `--max-experiments` | Max iterations | `50` |
| `--branch` | Git branch for winning commits | `optimize/karpathy-loop` |
| `--editable` | Paths the agent may edit (informational; enforced via prompt) | `src/` |

Results land in `logs/experiments.jsonl` inside the workspace. See [`demos/06_karpathy_optimization/`](demos/06_karpathy_optimization/) for a ready-to-run example and [`docs/EVALS.md`](docs/EVALS.md) for eval methodology.

Example excerpt:

Configuration reference

Copy .env.example to .env and set the API key for your chosen backend. Before each run, a pre-flight check validates configured models. Agent→model mapping and guardrail behavior are documented in docs/AGENTS.md and docs/GUARDRAILS.md.

Models by environment

Model IDs are in openrouter/<provider>/<model> form (see src/ai_team/config/models.py). Set AI_TEAM_ENV to dev, test, or prod to choose a tier.

Embeddings (crew memory) use OPENROUTER_EMBEDDING_MODEL (default: openai/text-embedding-3-small). Current IDs and pricing: OpenRouter models.

CLI options

The CLI has two top-level subcommands: run (build a project) and optimize (AutoOptimizer Loop).

```bash

Run with Claude Agent SDK

poetry run ai-team --backend claude-agent-sdk --team backend-api "Create a REST API for a todo list" ```

For step-by-step setup and troubleshooting, see docs/GETTING_STARTED.md.

Web Dashboard (FastAPI + React)

A production-grade browser UI with real-time WebSocket streaming, GitHub-dark theme, and side-by-side backend comparison.

```bash

Why compare?

The multi-agent framework landscape is moving fast. CrewAI, LangGraph, Claude Agent SDK, AutoGen, AWS Bedrock Agents — each makes different trade-offs around orchestration control, state management, human-in-the-loop, persistence, streaming, and cost. Rather than pick one and hope, this project runs the same team through multiple backends and lets the data decide.

Backend comparison

Run the same demo through multiple backends and compare:

python scripts/compare_backends.py demos/01_hello_world --env dev
python scripts/compare_backends.py demos/01_hello_world --env dev --with-claude

Produces a side-by-side report: output quality, cost, latency, token usage, error rate. Use --with-claude to include the Claude Agent SDK (requires ANTHROPIC_API_KEY).