The Agentic Tool Loop

Upgrade the default, never downgrade the specialist

The in-engine tool loop exists to give a plain, single-completion runtime the same layered capability a good coding harness already has natively — the role's manifest, a universal knowledge-worker floor (search, browser, Brain, data, compute), any bridged MCP tools, and the full agentis.* platform surface. It's explicitly never meant to downgrade an agent that's already bound to a powerful coding harness with its own superior loop — that agent keeps using its own.

One typed activity stream, every step

Every step — thinking, tool_call, tool_result — streams as a typed activity attributed to the specific node that produced it, so the run monitor, the triage view, and the canvas all read from one live feed instead of three surfaces that could drift out of sync with each other.

Continue