Agents & APIs: Paris Developer Meetup
Postman, with GitHub and Neo4j








- What I kept
- Our work habits are really changing, even at GitHub. Reviewing every line of AI-generated code by hand is becoming wasted time, provided the work is tightly structured and the checks are automated. For a long time we believed the quality of our tools made the quality of the result; the capabilities of harness providers now let us shrink our set of tools.
- What it changed
- It partly confirms the bet behind pythagoras-sec: fewer prebuilt tools, with connectors generated on demand.
Evening meetup on agents and APIs run by Postman, hosted at Microsoft in Paris, with three talks.
The talks
- From Autocomplete to Autopilot, Tugdual Grall (GitHub). Copilot’s path from completion to chat (February 2023) to agent mode (February 2025), then context engineering: instruction files, AGENTS.md, MCP, custom agents, skills and hooks. Every Copilot surface, CLI, app and IDEs, now runs on the same agentic harness exposed through the Copilot SDK, which routes models and tools, manages context and memory, enforces limits and recovers from failures. The Copilot CLI team ships 500 pull requests a week with 10 engineers.
- From Declassified Documents to Knowledge Graph Agents, Zaid Zaim (Neo4j). Public UFO records turned into a property graph (documents, witnesses, sightings, locations, with provenance and uncertainty), then connected to an agent. The retrieval agent picks its tool per question: vector and graph retrieval, Text2Cypher, Cypher templates written by domain experts, or an external source reached over MCP.
- From click to terminal: API engineering for agents, Anthony Viard (Postman). Designing tools for the agent itself, not only for the developer driving it: Postman’s CLI, plugin and skills in the terminal, used by Claude Code to build an API from scratch and to gate CI/CD on whether the API stays understandable to an agent.