Agnes Harness runs AI agents inside a customer's own systems with approvals

A customer's own systems come with their own data, their own approval rules and their own interfaces, and dropping an AI agent into that stack normally means forking the agent for every customer. Agnes Harness is an Apache-2.0 open-source agent runtime that AgnesAI Labs built to handle that work in one place. The repository is 26 days old, carries 774 GitHub stars and is still being committed to.

The README describes four parts in one process. A language model does the reasoning. A smaller decision model called Jev is scoped for structured routing and scoring choices, and the README says this piece is still being integrated. A harness holds session history, task state and reusable methods packaged as Skills. A separate device layer called MHS, short for Model Hardware Standard, lets a running task read sensors and request physical actions through WebSocket. The project calls itself a developer preview and pre-alpha, so a reader runs it from source rather than installing a release.

One session across the terminal and the browser

The CLI, the Web workbench and the SDK share the same backend sessions. An analyst can start a task in a browser, resume it from a terminal with /resume <id>, and the history, tool records and results carry across. Whatever the engineer adds in the terminal shows up back in the Web thread.

Every shell command the agent proposes waits for the operator by default: allow once, allow for the session, or deny. The Trace view records a timeline of model calls, tool calls and approvals, step by step. Plugins carry a version, a source, an integrity digest, a capability hash and a license, which the workbench shows before enabling anything. The FAQ states that ordinary backend plugins run as trusted in-process code, so the approval and sandbox cover the supported execution paths; a plugin itself can run whatever its code says, which is why installing an unfamiliar one is the operator's own risk.

A number the README reports itself

The project's README says Agnes Harness with Agnes 2.5 Pro Beta reaches a 21.7 percent overall pass rate and a 42.7 overall score on the public Agents' Last Exam leaderboard, the UC Berkeley RDI benchmark that scores complete agent systems on 1,500 or more expert-sourced tasks across 55 occupations. The leaderboard itself renders from JavaScript, so a reader has to open it to confirm the figure, and the README adds that results depend on model versions, settings and tool configuration. The project ships eleven virtual devices and a Mars World browser demo in which an agent coordinates a hopper, a robot and an airlock, so a reader can try the device side without any hardware.

For an engineer dropping an agent into a customer's environment, the trade is giving up a packaged installer in exchange for an audit trail the customer can read. Every tool call the agent made and every command the operator approved is in one place, with a version-locked plugin graph beside it. That changes what a post-release review is about, from "did it work" to "which step did you not expect".

Source

AgnesAI-Labs/agnes-harness on GitHub and the Agents' Last Exam leaderboard.