Flight Market Agents

A small demo: fan out N concurrent "agents," each pricing the same flight route in a different market's currency, then aggregate and compare. Built to show async orchestration (fan-out, concurrent I/O, streaming partial results) more than to be a real product.

What this actually does (and doesn't)

Each agent calls the Amadeus Flight Offers Search API for the identical origin/destination/date, but requests the price in a different currency (currencyCode=JPY, GBP, INR, etc.), then converts that back to USD so the numbers are comparable. That's a real, legitimate effect — currency-of-sale pricing genuinely doesn't track FX 1:1 — but it is not the "same flight, cheaper if you pretend to search from Japan" consumer-personalization effect you see on sites like Google Flights or Skyscanner. That effect lives in those sites' front-end revenue management and would require automated, high-volume querying against sites whose terms of service prohibit exactly that — not something this project does or is meant to do.

No proxies, VPNs, or IP/country spoofing are involved anywhere in this code. Every "market" here is just a currency parameter on a normal, licensed API call.

Two market-data modes

• Mock mode (default) — USE_MOCK_DATA=true in .env. Generates plausible fake fares locally with a staggered artificial delay per market, so you can run and demo the whole concurrent-agent flow instantly with zero setup. Every mocked result is labeled "source": "mock" in the API response and the UI.
• Live mode — USE_MOCK_DATA=false plus real Amadeus credentials. Calls the real API.

Two orchestration modes -- this is the part worth reading

The UI has two buttons, and they demonstrate genuinely different things:

• "Run concurrent agents" (app/orchestrator.py, POST /api/compare and GET /api/compare/stream) — a fixed async fan-out. It always queries all six markets, every time, via asyncio.gather / asyncio.as_completed. This is "agent" in the older, pre-LLM sense: concurrent workers, no reasoning involved. Good concurrency plumbing, not an AI agent.

• "Run AI agent" (app/agent.py, GET /api/agent/compare/stream) — a real tool-use loop against Claude. The model is given one tool, search_market_price (which just calls the same run_market_agent() underneath), and decides for itself which markets are worth checking for the given route, calls several in parallel when it wants to, and writes the final recommendation in its own words -- rather than you reading the lowest number off a board. Watch the board on a run: some rows may stay marked "Skipped" because the model chose not to check them. That's the actual routing/tool-use decision the job postings in this space are asking about.

This mode requires a real ANTHROPIC_API_KEY in .env (get one at console.anthropic.com) and makes real, billed calls to the Claude API -- small cost per run, but not free like mock market data. AGENT_MODEL defaults to claude-sonnet-5; swap it for whatever model you want to demo against.

The loop itself (run_pricing_agent in app/agent.py) is the standard Anthropic tool-use pattern: call the model, check stop_reason / tool_use blocks, execute whatever tools it asked for (concurrently, via asyncio.gather, reusing the same underlying market-search function as the non-agent mode), feed the results back as tool_result blocks, and repeat until the model stops calling tools and just answers. Capped at MAX_AGENT_STEPS = 4 so a confused run can't loop forever.

Setup

cd flight-market-agents
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
# .env already works as-is in mock mode - no keys needed to try it out.
uvicorn app.main:app --reload --port 8050

Then open http://127.0.0.1:8050.

Going live with real fares

1. Register for free self-service keys at developers.amadeus.com/register.
2. Put your client_id / client_secret in .env and set USE_MOCK_DATA=false.
3. Heads up: Amadeus's test environment (test.api.amadeus.com) runs on limited, static cached data — it often won't have offers for whatever route/date you pick, so you'll see "error": "No offers returned for this market" on some or all agents. Amadeus's own self-service tier auto-upgrades to the production base URL (api.amadeus.com) with a real free monthly quota (2,000 free Flight Offers Search calls/month at last check) — switch AMADEUS_BASE_URL to that once your app is approved for it, and use near-term real dates on well-traveled routes for the best chance of live results.

Where the concurrency actually shows up

app/orchestrator.py has two entry points:

• compare_all_markets() — asyncio.gather over all agents, returns once everything's done. Used by POST /api/compare.
• stream_market_agents() — asyncio.as_completed over the same tasks, yielding each result the instant it finishes as a Server-Sent Event. Used by GET /api/compare/stream, which is what the UI actually calls — watch the board and you'll see rows settle out of order, which is the visible proof the agents are running in parallel rather than one after another.

FX rates

app/markets.py has static, approximate FX rates for demo purposes. Swap in a live FX API (e.g. open.er-api.com, exchangerate.host) before relying on these numbers for anything real.

Deploy

Follows the same pattern as the other apps in this directory — see ../DEPLOYMENT.md. Port 8050, service name flight-market-agents, intended hostname flights.evancooperman.com. deploy/flight-market-agents.service and .github/workflows/deploy.yml are already set up.

Unlike the other apps here, this one has no database, but it does have secrets in .env — which is gitignored, so it never travels with git pull. The droplet needs its own .env created once, by hand, and left alone; ordinary deploys (git pull + pip install) don't touch it.

One-time droplet setup

1. GitHub secrets — on the flight-market-agents repo: DO_HOST, DO_USER=deploy, DO_SSH_KEY (same values as your other app repos).

2. Sudoers — add this app's service to the deploy user's narrow restart-only sudo rights:

   sudo visudo -f /etc/sudoers.d/deploy-restart

   Add flight-market-agents to the existing line, e.g.:

   deploy ALL=(root) NOPASSWD: /bin/systemctl restart time-management, /bin/systemctl restart social-planning, /bin/systemctl restart gifts, /bin/systemctl restart resume, /bin/systemctl restart flight-market-agents
   

   Then verify: sudo visudo -c.

3. Clone + venv + secrets:

   sudo -iu deploy
   cd /opt/apps
   git clone https://github.com/<your-github-user>/flight-market-agents.git
   cd flight-market-agents
   python3 -m venv venv && source venv/bin/activate
   pip install -r requirements.txt
   cp .env.example .env
   vi .env   # fill in real values - USE_MOCK_DATA=false and/or real API keys, as wanted
   exit      # back to your own sudo-capable user

   Mock mode (USE_MOCK_DATA=true, the .env.example default) needs no real keys at all if you'd rather stand it up first and add live credentials later.

4. systemd unit:

   sudo cp /opt/apps/flight-market-agents/deploy/flight-market-agents.service /etc/systemd/system/
   sudo systemctl daemon-reload
   sudo systemctl enable --now flight-market-agents
   curl -m 5 http://127.0.0.1:8050   # sanity check before wiring up the tunnel

5. cloudflared ingress — add to /etc/cloudflared/config.yml, above the catch-all - service: http_status:404 line:

     - hostname: flights.evancooperman.com
       service: http://localhost:8050

   Then route DNS and restart:

   sudo cloudflared tunnel route dns home-apps flights.evancooperman.com
   sudo systemctl restart cloudflared

6. Access — nothing to do if you're still on the wildcard Access application (*.evancooperman.com); it already covers this hostname.

7. Cache Rule (bypass) — Caching → Cache Rules → add flights.evancooperman.com to the existing bypass rule (or create one).

8. Push to main — CI takes it from here on future updates. Since this repo has no .git yet, that's git init, add the GitHub remote, then push, same as you did for gifts.