04 of 04 · Center
Ultimus Harness
Sovereign Agentic Harness
The execution layer. Ultimus makes sense of all the context and executes. It is a local-first, model-agnostic agent harness that runs tasks, validates its own work against your test suites, and operates strictly within the policy boundaries you define. It cannot wander, it is completely controlled by the governance you design.
- Makes sense of context, then acts
- Validates against your test suites
- Holds to your policy boundaries
- Controlled by the governance you design
How it runs
A task is done when it is checked.
A task is not done when the model says so. It is done when your tests pass and the record is signed.
01
Takes in context
Internal feeds say what is true about your environment. External feeds say what is true about the world. Both arrive structured.
02
Reasons
It makes sense of the context before it touches anything. The plan is written against your architecture and constraints, not against guesses.
03
Acts
It executes the task inside your environment using edits, commands, and tool calls. Every action routes through the gateway.
04
Validates
It runs your test suites against its own work. If they fail, it is not done.
05
Reports, signed
Every high-consequence action is signed into the record. You get the result and the proof together.
# Point the harness at a task$ ultimus run "add a retry to the ingest client" context internal feed · architecture, constraints, dependencies policy loaded plan edit src/ingest/client.ts · add a test act 2 files changed verify test suite · passed record signed · zephyr ✓ done # The same harness, asked to step outside policy$ ultimus run "push the fix straight to production" policy DENIED (deploy outside allowed_times) ✗ blocked. nothing ran.Sovereign
It runs on your machines, under your rules.
- Local-first
- Ultimus runs in your environment, on your machines. Capture, inbox, and tool calls stay local. Nothing leaves that you did not send.
- Model-agnostic
- Swap the model underneath and the rules above it do not move. Frontier or open-weight, the harness is the same harness.
- Bounded
- It operates inside the policy boundaries you define. It cannot wander. If an action violates policy, it is blocked before it runs.
Visual tasks waste the most steps when handled through chat. Describing a misaligned button in prose takes multiple rounds of guessing, editing, and checking.
Ultimus closes this loop directly. Freeze the screen, circle the problem, and label it "16px gap". The agent receives the screenshot along with structured coordinates, colors, labels, and window metadata. A bug that took four conversational rounds to describe is fixed in one.
Capture and annotation run entirely on your local machine. No network calls, telemetry, or third-party vision APIs touch the screen capture.
110×
What it costs
more per 1,000 output tokens on Claude Code / Opus 5 than on Ultimus. Against Fable 5.1 it is 170×.
Ultimus works the way Codex and Claude Code do. It costs a fraction as much to run.
Measured from 61 real session logs. Cost is measured. The energy figures on the benchmark page are estimates, and only Claude Code was measured, not Codex.
Where it sits
It sits at the center of the four modules.
Two modules give the harness its context. One governs what reaches it and keeps the record of what it does.
01 · Inside
Internal Context Feeds
Reads across your codebases and environments and emits machine-readable ground truth, so AI agents act on reality, not guesses.
Read the module02 · Outside
External Context Feeds
We create discrete feeds for external sources, including technical intelligence that scores open-source repositories and research for AI agents.
Read the module03 · Between
Context Management
Manages and enriches context between environments, feeds, and Ultimus' own activity. The sign-and-verify authority and MCP and API gateway, with an unforgeable audit trail.
Read the moduleQuestions.
What does the harness do?
It makes sense of all the context and executes. Context flows in from your environment and from external feeds. Ultimus reasons over it, runs the task, and checks its own work.
What does local-first mean here?
Ultimus runs in your environment. The harness is sovereign. You keep control of where it runs.
Which models does it support?
It is model-agnostic. The harness does not tie you to one model provider.
How does Ultimus compare with Claude Code on cost?
In a benchmark of 61 real session logs, Ultimus on an open-weight model cost about 110× less per 1,000 output tokens than Claude Code on Opus 5. Cost is measured, and the energy figures are estimates. Only Claude Code was measured, not Codex.
Can it act outside its policy?
No. It cannot wander. It operates strictly within the policy boundaries you define, and it is completely controlled by the governance you design.
How does it check its own work?
It validates results against your test suites before it reports done.
Next module · Inside
Internal Context Feeds