What you should be able to do
Build agent loops with tool schemas, guardrails, tracing, handoffs, memory, and clear failure boundaries.
CapstoneCreate a research agent that gathers sources, drafts a plan, and records traceable decisions.
Models that plan, use tools, and recover
Agents
An agent is an AI system that can inspect state, decide what to do next, call tools, and continue until a goal or stop condition is reached.
Observe
Plan
Call tool
Check result
Stop or continue
Concepts to master
- Tool calling and schemas
- Planning loops and stop conditions
- Handoffs between specialized agents
- Tracing and observability
- Safety boundaries for autonomous actions
Builder path
How to turn this topic into a working project.
Use this as the bridge from reading to implementation. The goal is to build a small, inspectable version before adding frameworks or production complexity.
- Open VS Code and create a small TypeScript or Python project with one read-only tool, such as `search_docs` or `get_order_status`.
- Define a strict schema for that tool, validate model-requested arguments, and return structured observations.
- Wrap the model call in a step-limited loop that can call tools, record traces, handle tool errors, and stop with a final answer.
- Add human approval for any write action before you add more tools, memory, handoffs, or multi-agent orchestration.
Primary sources
Back to roadmap Open first source