BlueHomer
Navi: an AI agent you can safely let act
Most AI assistants can talk. Navi can act — and it does so under the same rules a permitted employee would follow.
What Navi is
Navi is BlueHomer’s execution brain. Given a request in plain language, it discovers what the platform can do, plans a sequence of steps, asks an AI model which tools to use, then executes those tools under the acting person’s own identity — checking authorization at every step, recording each action, and streaming live progress back to the user.
The distinction that matters is not fluency. It is that an agent which *acts* touches real systems, and therefore has to inherit the same controls a person does: identity, permission, approval, audit, and a way to stop it.
It composes governed building blocks — it never invents access
Navi holds no model keys and defines no tools of its own. It orchestrates registered, typed capabilities from a governed catalog, which is the only path to action on the platform. Adding a new capability means registering an action — not giving the AI a bigger blank check.
| Approach | What the agent holds | Blast radius when it is wrong |
|---|---|---|
| Agent with credentials | Database or API credentials, free-form query ability | Anything those credentials can reach |
| Agent with a tool list | A fixed set of functions, often unscoped to the user | Whatever the tools allow, for any user |
| Navi | No keys. A registry of typed capabilities, executed as the acting user | Exactly what that user was already permitted to do |
Autonomy without losing control
Navi operates in selectable modes — from "confirm before every change", to "run reads automatically, confirm writes", to more autonomous operation. Crucially, autonomy removes conversational prompts, never governance. Privileged or destructive actions still pause for human approval, every time.
- You ask for an outcome in plain language.
- Navi discovers the capabilities available to *you* and plans a sequence.
- Each step is authorized against policy before it runs — not after.
- Reads proceed; privileged or destructive steps stop and wait for a human decision.
- Every tool call, result and produced artifact renders as its own inspectable entry, live.
- A step budget bounds the run, so a flawed plan cannot run away with your resources.
Total transparency, and a stop button
- Every step Navi takes is streamed and inspectable — the opposite of an opaque black box.
- You can cancel at any moment, mid-run.
- Every action lands in the tamper-evident audit trail, attributable to the person on whose behalf it ran.
- A built-in step budget is a containment boundary, not a billing feature.
Ambient, not just conversational
Beyond chat, Navi can be triggered on a schedule, by platform events, or when new files arrive — enabling proactive, governed automation. The same authorization and audit rules apply whether a human or an event started the run.
Frequently asked questions
How do you audit what an AI agent did on your behalf?
Every step Navi takes — each tool call, each result, each produced artifact — is recorded in an append-only, hash-chained audit trail, attributed to the identity that authorized it. The record is exportable as evidence and its integrity can be independently verified.
Can the agent act without approval?
Reads can run automatically in the higher autonomy modes. Privileged or destructive actions always pause for human approval regardless of mode — autonomy changes how much the agent asks you conversationally, not what governance requires.
What stops a runaway plan?
A step budget bounds every run, approval gates stop privileged actions, and the run can be cancelled mid-flight. If governance cannot be verified for a step, the platform fails closed and the step does not execute.
Does the agent see data it should not?
No. Navi executes under the acting user’s identity, so field-level masking and row-level filtering apply to the agent exactly as they apply to that person. It cannot see more than the user could see themselves.