Agents and Tool Use MOC
Agents are model-driven systems that act over time: calling tools, reading environments, writing code, navigating browsers, using memory, and recovering from mistakes.
Why It Matters
Frontier labs increasingly care about:
- Coding agents.
- Browser/computer-use agents.
- Research assistants.
- Tool-using assistants.
- Long-horizon task completion.
- Verifiable work in external environments.
Text prediction becomes more valuable when connected to action and feedback.
Core Ideas
Tool Use
The model chooses structured actions:
- Search.
- Calculator.
- Code execution.
- Database query.
- File edit.
- Browser click.
- API call.
Primary source:
Reasoning + Acting
The model alternates between reasoning traces and actions in an environment.
Primary source:
Memory
Agents need memory beyond one response:
- Short-term context.
- Long-term notes.
- Retrieved documents.
- Task state.
- User preferences.
- Tool outputs.
Verification
Good agents use external checks:
- Tests.
- Type checking.
- Unit test generation.
- Program execution.
- Search validation.
- Human confirmation for risky actions.
Environments
Agent ability depends on environment design:
- Are actions observable?
- Are errors recoverable?
- Is reward/checking clear?
- Can the agent inspect state?
Starter Project
Build a coding-agent eval loop:
- Give the agent a small repo bug.
- Let it inspect files, edit, and run tests.
- Log every action.
- Compare with a no-tool baseline.
- Track pass/fail and error recovery.
Deliverable:
- A mini report on where the agent fails: planning, search, editing, testing, or recovery.
Papers / Resources
What Good Looks Like
You can:
- Design tool schemas.
- Build an agent loop.
- Add verification.
- Evaluate long-horizon tasks.
- Distinguish model failure from scaffold failure.
- Recognize when "agent" is just an overcomplicated prompt.