Agents and Tools

An agent is an LLM that runs in a loop, using tools to act on the world and deciding its own next step until a task is done. That autonomy is powerful and risky — so the discipline here is as much about control as capability.

Anthropic's core advice: don't build an agent if a workflow will do. Start simple, add autonomy only when the steps genuinely can't be predetermined, and keep the system transparent.

Mental model

An agent is a model embedded in a control loop with observations, tools, state, budgets, and stop conditions. Capability comes from the loop; safety comes from treating model output as untrusted proposals and keeping authority in deterministic infrastructure.

Roadmap: interfaces to autonomy

• Workflows vs agents: when to use which
• Tool & function calling
• Designing the agent-tool interface
• The Model Context Protocol (MCP)

How agents work

• The ReAct loop: reason + act
• Planning & decomposition
• Agent memory
• Multi-agent systems & handoffs

Keeping them safe & working

• Guardrails & human-in-the-loop
• Autonomy & least privilege
• Agent failure modes
• Evaluating agents

Connects to: Context Engineering · AI Safety and Security · Evaluating Agent Systems

Core sources

• Anthropic — Building Effective Agents (Dec 2024) — the workflow/agent taxonomy and "simplest thing that works" doctrine; the branch's backbone.
• Anthropic — Writing effective tools for agents (2025) + How we built our multi-agent research system (2025) — measured tool-design guidance and production multi-agent economics.
• MCP specification — the open standard for model↔tool integration.
• Yao et al. 2022 — ReAct (arXiv:2210.03629) — the loop every agent still runs.
• OpenAI — A Practical Guide to Building Agents (2025) — the other vendor's converging playbook; useful for calibration.