01 · The problem
Built five successive generations of the same planner, linear, parallel, orchestrator, human-in-the-loop and an MCP factory graph, keeping every generation runnable so the architectural trade-offs can be compared rather than argued about.
02 · How it works
- 01
Served all five concurrently through Aegra on AWS Bedrock with a standalone MCP server exposing 10 travel tools, and covered the backend with 630 tests against a real Postgres.
03 · What it cost, and what it returned
Wired live Playwright browser automation the user can take over mid-run, so the agent hands control back at exactly the point where a human should decide.
