AI News Feed
Market watch
Amazon Says PHP2Shell Attacks Began Four Days Before Patches as Attempts Reach MillionsStrava Revamps Route Discovery With Smarter Search and Offline MapsSpotify confirms QR code checks for Family plan sharingFortune report puts next Xbox at $1,200 to $1,800, testing Microsoft's affordability pledgeGoogle Funds Virtual Power Plant to Free Grid Capacity for Data CentersGaelic Horror Series Wolf Bay to Stream Free From October 23
Amazon Says PHP2Shell Attacks Began Four Days Before Patches as Attempts Reach MillionsStrava Revamps Route Discovery With Smarter Search and Offline MapsSpotify confirms QR code checks for Family plan sharingFortune report puts next Xbox at $1,200 to $1,800, testing Microsoft's affordability pledgeGoogle Funds Virtual Power Plant to Free Grid Capacity for Data CentersGaelic Horror Series Wolf Bay to Stream Free From October 23
Products & Applications

Open-source CATT 0.2.1 tests services against Muse/Instinct-like agents in CI

CATT 0.2.1 lets teams test services against surrogate Muse or Instinct-like agents locally or in CI.

AI News Feed · 2026-10-23 09:46

The team said it built CATT to reduce its own frustrations with building an MCP or API for these agents to consume. Testing against such tools is hard, the post said, because agent memory and account statefulness make reproducibility difficult in the ways needed for repeatable and reliable outcomes between agents and the APIs or MCPs they interact with.

CATT addresses that by letting developers test their services against a surrogate Muse or Instinct agent. The tool is not affiliated with either company, the post said. Testers define starting conditions, approvals and success conditions for their service's state. CATT then records the full interaction, including tool calls and responses, so it can diff what happened from what the agent reports. Runs can be executed in local builds or through CI, with or without LLM backing.

CATT is designed to emulate important parts of personal-assistant interactions, including tool calls and command-line tool access, simulated approvals for service actions, outcomes tracking and comparison, and an optional MCP profile based on an unofficial Muse template.

The developers acknowledged limits to how much they know about assistant-specific behavior. The tool labels its assumptions as documented, unofficial, assumed or unknown for developer review, and the team said it will do its best to keep the tool updated as new insights about the internal workings of these assistants come to light.

It described the area as active research and cited a recent report on how Instinct uses Exa for web search as an example.

The post asked people building connectors or MCPs to get in touch and to describe cases where assistants behaved incoherently, used stale information or proclaimed success without evidence. It also asked what other tests should be added. For getting started, the developers recommend pointing a coding assistant at the repository and asking it to integrate CATT with a CI or local testing pipeline.

Editor's Summary

CATT 0.2.1 is an open-source tool for testing services against surrogate Muse- or Instinct-like agents, with interaction recording, outcome diffing and CI support. It labels assumptions about assistant behavior and asks MCP and connector developers for feedback on failure cases. The release targets reproducibility problems caused by agent memory and account state.