Guides

Task-shaped instructions for each step of the loop. Each guide assumes you have a project (science-adk init) and ends with something you can verify.

• Write an agent node

  ---

  The programming model in practice: ports, execute, inputs, tools, logs, artifacts, and the mistakes validate will catch.

• Write and register tools

  ---

  Turn a Python function into a tool, get the schema for free, and fail loudly instead of returning a sentinel.

• Connect MCP servers

  ---

  Stdio subprocesses and remote HTTP servers, with tokens read from the environment.

• Validate a workflow

  ---

  Read the report, fix the common errors, and know which warnings are worth arguing with.

• Run and read a trace

  ---

  --only, --reuse, run directories, and how to answer "what actually happened" from trace.json.

• Audit, report, campaign

  ---

  Turn a green run into a recorded finding with a defensible rationale.

• Evolve an experiment

  ---

  Lineage, learnings, and deciding whether to optimise the current hypothesis or replace it.

The order they are usually needed

graph LR
    Tools[write tools] --> Agents[write agents]
    Agents --> Validate[validate]
    Validate --> Run[run]
    Run --> Audit[audit]
    Audit --> Evolve[evolve]
    Evolve --> Agents

If a step is failing rather than missing, the debugging skill is the faster route.