Harnesses & Deep Agents

In one line

A harness is the scaffolding around the model. Deep Agents is a ready-made one.

Key idea

2026 lesson: a great model fails without a good harness. Deep Agents bundles:

• write_todos planning tool — keeps focus on long tasks.
• virtual filesystem — ls, read_file, write_file, edit_file.
• sub-agents via a task tool — fresh context, returns one summary.
• a tuned main loop.

from deepagents import create_deep_agent
agent = create_deep_agent(model="claude-sonnet-4-6", tools=[search], system_prompt="Plan, then work in steps.")

New in June 2026

• Rubrics. RubricMiddleware (beta) makes the agent grade its own output and keep fixing it until a checklist passes. See Middleware for the setup. Docs: rubric.
• Programmatic subagents. You can now launch subagents from inside interpreter code, not just the task tool — good for fanning out work and merging results. Docs: programmatic subagents.

Links

• Deep Agents overview
• deepagents repo
• Anatomy of an agent harness
• Full guide — appendix G (Deep Agent)

My notes