Built agent-first

Alloy was built from day one for an AI agent to do the investigation. Why that matters, and how it differs from a tool with an agent added.

Alloy was built from day one for an AI agent to do the engineering work on your robot data. It is not a viewer with an agent added, and it is not a general coding agent you have to build a harness around.

Everything in Alloy exists so the agent can finish the job: the data layer, the tools, the knowledge base and the integrations. A person asks the question. The agent finds the runs, reads the evidence, works out the likely cause and leaves something your team can act on.

Judge an agent by what it leaves behind

Many tools now have an AI agent. The useful test is what you have when the agent finishes.

	An agent added to a tool	Alloy, built agent-first
Where the agent works	Inside one tool, on the data that tool shows	Across your robot data, tickets, code, telemetry and docs
What a run leaves behind	A layout, a script, a tagged event	A likely cause, each finding linked to the moment in the recording, a next step and a ticket if you want one
What happens next time	The next person starts again	The knowledge base keeps the answer, so a known issue gets a known answer
When it runs	When a person opens the tool	On every new run and every new ticket, without a person starting it

What agent-first means in practice

The data is ready for the agent. Recordings sit in Mesh Storage or stay in your own store, indexed so the agent can query a whole fleet in one question. Nobody downloads a large log to a laptop.
The tools are built for robot data. The agent reads MCAP, ROS bags, flight logs and telemetry, aligns time across sensors, and runs analysis in a sandbox. It does not guess from a summary.
The agent shows its work. Every finding links to its evidence. When the data cannot confirm something, the agent says so. See Security and trust.
The agent works where your team works. It answers in Slack, on Jira tickets and in the app. It reads your Confluence runbooks and your GitHub code.
The agent works without being asked. Recurring reports and scenarios review every new run, so the robot that needs a person is found before the field reports it.
Your own agents can use it too. Claude, Codex and Cursor connect to Alloy through MCP and work with the same data and tools as your engineers. See Alloy for AI agents.

Why this is hard to add later

Adding an agent to an existing tool gives the agent that tool's view of the data: one recording, one panel, one layout at a time. Real robot investigations cross many runs, many sensors, the code and the ticket history. An agent that can only drive a viewer stops at the edge of the viewer.

Agent-first also means the hard parts are already solved: many recordings in one question, the code context next to the logs, memory of past cases, answers that cite their evidence, and models tested on robotics tasks. A team that adds an agent to a tool, or wires a general agent to its own data, meets each of these problems one at a time.

An analogy engineers know

Coding went the same way: the biggest gains came from tools built around the agent, where the agent does the work and the engineer reviews it. Alloy is built that way for robot data.

The proof is public

We test the models behind the agent on real robotics tasks, and we publish how frontier models perform inside the Alloy agent: the Robotics Engineering Benchmark, with score, cost and speed for every setting. How we choose models is on Model testing.