← back
Agents Building Agents - Alfonso Graziano, Nearform
Takeaway
Coding agents can improve other agents through repeated changes measured against explicit evaluation targets.
Summary
- Nearform builds golden datasets with domain experts, specifying expected answers, tool calls, parameters, and action sequences.
- A basic Mastra agent initially passes 18% of a toy evaluation suite because it lacks necessary tools and context.
- AutoAgent repeatedly runs evaluations, edits prompts and code, adds tools, and checks whether results improve.
- The toy example rises to 83% in about ten iterations; the speaker also reports a 10% improvement on some evaluations of an already optimized production agent.
autoagentagent-evalsmastra
Original description
Building an AI agent for a real team is not a prompt problem, it is a systems problem. In this session we walk through a practical, production-minded workflow for building an agent using a coding agent, and designing the codebase so that this loop stays reliable as complexity grows. The core pattern is two agents with different jobs. The coding agent is the builder: it writes and changes the agent’s codebase. The agent you are building is the product agent. It is the custom agent you ship for a client or for internal use. A key example is self-healing evals. We maintain an eval suite that exercises the product agent across representative tasks. When an eval fails, the builder agent runs the eval, inspects the failure artifacts, proposes a targeted fix to the correct layer (context, tool contract, or code), and opens a PR with a short report explaining what changed and what is still missing. If the agent cannot safely resolve the failure, it escalates by requesting specific human input and explaining exactly why it is blocked. Speakers: Alfonso Graziano (Nearform): Alfonso is a Software Engineer led by curiosity and passionate about new technologies LinkedIn: / alfonso-graziano GitHub: https://github.com/alfonsograziano