Agentic AI Starter: Plan -> Act -> Reflect

This is a complete starter project for a multi-agent research workflow:

• Planner agent creates tasks
• Worker agent collects evidence from web search and summarizes
• Critic agent checks gaps/contradictions and requests follow-up tasks
• Orchestrator returns a final brief with sources and next checks
• FastAPI exposes the workflow for apps/clients
• Expo mobile app lets you run the agent from your phone

Project structure

.
|- mobile-app/
|  |- App.tsx
|  |- app.json
|  `- package.json
|- src/agentic_research/
|  |- agents/
|  |  |- planner.py
|  |  |- worker.py
|  |  `- critic.py
|  |- tools/
|  |  |- web_search.py
|  |  `- memory.py
|  |- config.py
|  |- llm.py
|  |- models.py
|  |- orchestrator.py
|  |- prompts.py
|  |- api.py
|  `- cli.py
|- tests/
|  |- test_orchestrator_smoke.py
|  `- test_api_smoke.py
|- .env.example
`- pyproject.toml

Setup

1. Create virtual environment:

python -m venv .venv
.venv\Scripts\Activate.ps1

2. Install dependencies:

pip install -e ".[dev]"

3. Configure env:

Copy-Item .env.example .env

Then edit .env and add OPENAI_API_KEY.

Run CLI

agentic-research "Perplexity AI"

or

python -m agentic_research.cli "Perplexity AI"

Outputs are saved in output/ as both .json and .md.

Run backend API

agentic-research-api

or

python -m agentic_research.api

API endpoints:

• GET /health
• POST /research with JSON body:

{
  "topic": "Perplexity AI"
}

Run mobile app (Expo)

1. Start backend API in one terminal:

agentic-research-api

2. Start mobile app in another terminal:

cd mobile-app
npm install
npm run start

3. Open the app in Expo Go, then set API Base URL to your laptop LAN IP: http://192.168.x.x:8000

How the pattern works

1. Plan: PlannerAgent.plan(topic) creates prioritized tasks.
2. Act: WorkerAgent.execute(task, search_results) turns raw results into findings.
3. Reflect: CriticAgent.review(topic, findings) flags gaps and suggests follow-ups.
4. Synthesize: ResearchOrchestrator builds the final executive brief.

Extend ideas

• Add tool routing (news API, Crunchbase, SEC, docs)
• Add a verifier agent for citation checks
• Add retries and confidence-weighted ranking
• Add memory persistence (SQLite) for multi-run learning