Technical architecture
A local control plane for agents that have to move a real number.
Nimrobo is a macOS outcome harness. It keeps the goal, evidence, and operating policy on your machine, gives one bounded action to an AI agent, and feeds the measured result into the next run.
Runs on
macOS
State
Local-first
Isolation
Apple Seatbelt
Unit of work
One action per run
The control plane
The agent does the work. The harness decides what counts.
A language model can plan and call tools, but it cannot tell from its own output whether a business result improved. Nimrobo keeps that judgment outside the model. The north star, prediction, evidence window, and gate are registered before the result arrives.
- 01
Read the outcome
The harness loads one north star, its guardrails, the active levers, and the evidence accumulated by earlier runs.
- 02
Choose one bounded action
It explores an unproven hypothesis or exploits one that already held up, then hands exactly one action to the agent engine.
- 03
Act through tools
The agent reads files, runs commands, prepares changes, and stops at approval boundaries before consequential external actions.
- 04
Measure what happened
Shipped artifacts are tied to delayed reward from the real destination, not to a proxy score produced inside the model.
- 05
Keep the evidence
The result joins the local record, so the next run can use what worked, what failed, and where evidence is still missing.
Local-first state
The memory is durable because it is data, not another prompt.
Customer data, credentials, outcome history, and reward evidence stay on the Mac by default. Model requests go to the provider you connect; Nimrobo does not require a hosted project database to remember the loop.
Outcome folder
Goal, operating policy, gates, and compact records for each run.
Evidence store
Hypotheses, experiments, artifacts, reward snapshots, and verdicts.
Session history
The agent transcript and subagent tree behind the work.
Git history
A reviewable record of changes to the outcome and the surface it acts on.
The evidence graph
north star ← lever ← hypothesis ← experiment ← run
Explore runs attach evidence to an experiment. Once a hypothesis is supported, exploit runs attach yield directly to that known action. Project metrics answer whether the overall outcome moved; artifact reward answers what one shipped action earned. The two are never silently substituted for each other.
Execution boundaries
Isolation enforced by the operating system, approval controlled by you.
Sandboxed shell commands run under Apple's Seatbelt mechanism. A blocked file or network operation fails at the operating-system boundary. Approval mode is a separate control, so you can narrow both what a command can reach and how often the run has to stop.
Read about guardrails and approvalstandard
Network: BlockedOutcome folder and temporary files
Runs silently inside the narrow sandbox
elevated
Network: OpenApproved files and network
Asks before every elevated call
no-sandbox
Network: OpenFull machine access
Explicit approval on every call
Default
Routine sandboxed work proceeds; higher-risk calls still ask.
Manual
Every write, edit, and command waits for approval.
Plan
The agent can inspect and propose, but it cannot change the surface.
Agent engine
A swappable worker underneath a stable outcome loop.
The agent engine reads the bounded assignment, plans, calls tools, and reports what changed. The harness keeps ownership of the goal, the explore-or-exploit choice, the approval boundary, and the evidence that determines whether the action worked.
Keeping those layers separate means the model or agent runtime can improve without erasing the accumulated learning behind the outcome.
See the agent engine docsThe practical version
Give the agent a goal. Keep the controls and the evidence.
Start one outcome on your Mac, watch each action run, and let the next decision use the measured result instead of starting from a blank chat.
Download Nimrobo