Agents Stopped Calling Tools and Started Writing Code

The most interesting AI progress this week happened in the layer almost nobody photographs: how agents actually decide what to do.

Code as Actions

Hugging Face significantly pivoted its smolagents framework around a simple idea with big consequences. Instead of the agent emitting brittle JSON to invoke a predefined tool, the agent just writes Python and runs it.

• Core logic fits in roughly 1,000 lines of code
• Around 30% fewer execution steps and LLM calls versus schema-constrained approaches
• 67% success rate on their benchmark, beating traditional tool-call setups
• Python's full expressiveness available for complex logic instead of whatever the JSON schema anticipated

I converted one loop in my alerts bot from strict function calling to generated-Python execution as an experiment. Same tasks, fewer round trips, and the failure mode changed in an interesting way: instead of malformed JSON errors, I now get actual stack traces, which are dramatically easier to debug.

JSON tool calls made agents talk like machines. Letting them write code lets them think in their native language.

The Catch, Obviously

Arbitrary code execution is exactly what sandboxes were invented for. The moment your agent writes real code, your containment requirements jump a tier, which loops right back to last week's infrastructure stories. Nobody ships this pattern without a proper box around it.

Routers and Plain English

Two more releases rounded out the week. Cursor Router promises frontier-quality results at up to 60% lower cost by picking the cheapest model per task, treating models like interchangeable parts instead of a religion. And Cohere introduced North Automations, letting enterprises describe whole workflows in plain language with branching, approvals, versioning, and token-cost visibility built in.

• smolagents: agents write code, fewer steps, needs sandboxing
• Cursor Router: model selection as an optimization problem
• North Automations: workflow orchestration without engineering tickets
• Anthropic Managed Agents: effort levels now persist per agent across sessions

The harness layer is where the money gets saved now. The model matters less than what wraps it.

Cheers, Yassen