An agent harness for the messy part: the customer's site

What it does

Agnes Harness (AGH) is a TypeScript agent runtime aimed at Forward Deployed Engineering — the work of shipping AI into a customer’s actual environment, where the order system, the approvers, and the devices differ at every site. Business capabilities live in plugins (backend tools, MCP connections, Web panels, Skills), while the harness runs the loop, keeps task state, and records everything. CLI, Web, and SDK clients share the same sessions, so a task started in a browser can be resumed in a terminal with full history.

The interesting bit

The trust posture is the product: plugins show their source, integrity digest, capability hash, and license before enabling; commands wait for human approval by default; and a Trace view keeps a step-by-step timeline of model calls, tools, and sign-offs. There’s also MHS, AGH’s own device protocol (independent of MCP), which connects hardware over WebSocket so device reads and actions join the same task flow — with a simulated Mars base and eleven virtual devices to try without buying a robot.

Key highlights

• Shared sessions across CLI, Web, and SDK — resume anywhere, history intact
• Plugin governance: install, enable, update, roll back, with integrity verification
• Default-on command approvals and a full Trace timeline for audit questions
• MHS/MOS device protocol with a conformance suite (mhs-check) and AgnesHub as an optional plugin
• Skills capture task methods so the next deployment starts from the last one

Caveats

• Firmly pre-alpha: APIs, configuration, and plugin interfaces are all subject to change
• Not a sandbox for arbitrary plugin code — ordinary backend plugins run as trusted in-process code; approvals only cover supported execution paths
• The “Jev” cerebellum component is still in progress; the main path currently uses a built-in Core loop

Verdict

Worth a look if you deploy agents into customer environments and need approvals, traces, and reusable per-site plugins. If you wanted a finished, stable platform or a hosted service, this is explicitly neither.