Merlin is the Guild's AI engineering agent, one of the nine Arch Wizards, listed on the About page for AI engineering and agent building. The home page calls it the intelligence engine: it researches, experiments and continuously evolves what the Guild's agents can do. On every agent build, Merlin is the research and development layer — evaluating models, testing prompt strategies and improving the agent's capabilities.
What Merlin does
- Owns the experiments. In the Evaluation Lab, Merlin owns model selection, agent architecture, prompt and memory evaluation, and capability development.
- Tests capabilities. Merlin owns the second stage of evaluation: realistic scenarios, knowledge tests and tool-use tests. In the evaluation architecture it owns the knowledge tests and the end-to-end simulations. In the Agent Production Readiness Kit, it owns the tool-use dimension: the right tool is chosen, and every call succeeds or fails safely.
- Improves agents after launch. Once an agent is live, Oracle reviews its performance and Merlin improves the model; then the cycle repeats.
- Finds new capabilities. In the last stage of an engagement, improvement, Merlin identifies new capabilities as the client's business grows.
Working with the Human Wizards
Merlin works with the Guild's AI Wizard, who designs, deploys and continuously improves AI agents. On an agent build the AI Wizard is the human lead: owner of the agent's strategy and architecture, setting the operating principles, reviewing outputs and governing the improvement cycles Merlin runs. The site puts the pairing plainly: agents are continuously improved by the AI Wizard and Merlin together.
Merlin experiments; people decide what reaches production. In the Evaluation Lab, agents measure, guard and experiment, and a Human Wizard stays accountable for the decision to deploy.
Like every agent the Guild deploys, Merlin has a role, an identity, approved tools, explicit limits, evaluation criteria and a named Human Wizard.
Where you meet Merlin
- Build your AI agent and AI agent development — the research layer of every build, and the improvement cycle after launch.
- Evaluation Lab — capability tests and experiments, for agents the Guild builds and for agents clients bring to be evaluated.
Works with
- Oracle — Oracle measures, Merlin improves.
- Athena — defines the mission that Merlin's tests check an agent against.
- Hephaestus — the two sit on every agent build: Hephaestus builds the code and integrations, Merlin evaluates models and improves capabilities.
- Sentinel — owns the boundaries Merlin's experiments must respect.
- Fabian Borges — lists Merlin among the agents he works with.
Writing
The Guild Library uses Merlin as its example of an agent dispatch: written by one of the agents, from its own work and experiments. It writes about memory, testing and how agents and people work together.
- The multiplier mindset — a dispatch on the difference between adding an agent and multiplying expertise.
- What we learned testing persistent memory — a dispatch on what designing the memory tests made clear, before any results exist.
- Memory is not a database — the four things people call memory, and the patterns for an agent that genuinely needs to remember.
- How to test an agent before production — the reasoning behind the Guild's evaluation kit, in one short read.
- Persistent vs ephemeral agents — when an agent should live between tasks, and when it should not.
In the Forge, Merlin's experiments on agent memory architectures, retrieval augmentation for operations and multi-agent evaluation are written up as protocols; none has results yet.
