Agents that write their own tools

For an agent to be really useful, it needs tools — specialized pieces of code that help complete specific tasks, like browsing the web. Today, these tools are built manually by developers who try to anticipate what the agent might need. When they miss something, work stops until a developer notices and builds a new tool.

We propose a different approach: agents that write their own tools when they encounter new challenges.

How agents identify tool needs

For example, imagine an agent working on a weather forecasting project in Superagent. It might break this down into tasks like:

1. Import historical weather datasets
2. Wait for human to verify analysis assumptions
3. Build seasonal forecasting model
4. Generate report with findings

If the agent discovers it can't identify seasonal patterns with its current tools, it recognizes this gap and starts planning a new tool to fill it.

Writing new tools

Once the agent knows what tool it needs, it follows a test-driven development (TDD) approach:

First, the agent writes tests that define what the tool should do. After writing tests, the agent:

1. Implements the tool
2. Runs it in an isolated testing environment
3. Fixes the tool when tests fail
4. Repeats until all tests pass

The sandbox environment keeps the development process safe by preventing unintended interactions with other systems. We recommend E2B - an open-source runtime for executing AI-generated code in secure cloud sandboxes. Additionally, environment secrets can be managed here, ensuring sensitive data like API keys and credentials remain secure.

Tool evaluations

Tool evals focus on whether a tool effectively completes its intended task, and help agents improve at writing and using new tools. The evaluation process combines:

• Quantitative metrics (runtime, resource usage, error rates)
• Task completion (did it solve the problem it was built for?)
• User feedback (was the solution useful?)

These evaluations feed into a reinforcement fine-tuning process where human feedback guides the agent's learning. The agent trains a reward model from human evaluations of tool performance, uses this model to optimize its tool-writing capabilities, and continuously improves as new evaluations come in.

Effects on human-AI collaboration

Current agent implementations require humans to constantly anticipate and provide for agent needs, creating a dependency that limits both human and agent potential. Agents that write their own tools fundamentally transform this relationship. Instead of spending time developing basic tools, human developers can focus on strategic oversight and direction-setting.

This shift transforms how organizations work with AI in several important ways. Humans can now focus on providing high-level guidance rather than handling tactical implementation details. Organizations can scale their AI capabilities without needing to proportionally increase their development resources. Perhaps most significantly, there's a compound learning effect as tools build upon tools, accelerating the development of new capabilities.

As agents become more proficient at tool creation, the human role evolves from tool provider to strategic evaluator and guide. This evolution not only improves efficiency but enables organizations to tackle increasingly complex challenges by leveraging the growing capabilities of their AI systems.