# The LLM rewrote the email after I approved it

> Source: <https://www.crusbro.com/en/product-architecture.html>
> Published: 2026-08-21 07:11:02+00:00

# crusbro agent-OS

Agent Operating System · HARNESS · 600,000+ hours

Agent failures are rarely “the model isn’t smart enough.” They are boundary failures: outbound sends, writes, session drift, version drift. crusbro agent-OS hardens harness insight into a runtime—so models finish defined goals under gates, freeze and replay.

## What it is

crusbro agent-OS is a mature harness for driving models to complete defined work: goals, tools, memory, failure handling, human gates and evidence in one foundation. The model reasons; agent-OS keeps that reasoning **on an acceptable business track**.

What to finish · How to accept

**→**

Harness · Gate · Freeze · Replay

**→**

System actions · Documents · Devices · Reports

## What is not substitutable

Anyone can rewrite another agent loop. What is hard to replace: production insight into harness failure modes—and defaults shaped by runtime data.

| Attribute | Why generic frameworks fall short |
|---|---|
| Harness engineering insight | Addresses tool hallucination, accidental outbound, session confusion, missing rules and version drift with gate + freeze + replay—not a demo loop. |
| 600,000 hours of runtime | Failures and dispositions feed defaults: retries, idempotency, gate thresholds and risk tags come from real operations. |
| Freeze before outbound / write | Immutable snapshot before high-risk actions; execute only after approval. Models cannot bypass the physical gate. |
| Order-level evidence packs | Sessions, tool calls, approvals and outbound tied to business order IDs. Acceptance unit is business fact, not scattered logs. |
| Config revision bound to Run | On failure, return to the exact rules, prompts and tool permissions then in force—with canary and rollback. |
| One acceptance language across lines | Automation, drawing, vision and robot share orchestration and audit. Org capability transfers; no new platform per agent. |

## Role in build and orchestration

Teams write goals, tools, gates and acceptance. agent-OS turns proven harness behavior into configurable defaults.

| Stage | What the team does | What agent-OS hardens |
|---|---|---|
| Define | Goals, boundaries, acceptance | Task state machine, pass/fail, milestone gates |
| Connect | Mail, forms, MES/QMS, device actions | Skills / Tools constraints, idempotency keys, risk tags |
| Gate | Who approves, when humans must step in | HITL, freeze snapshots, high-risk blocks |
| Pilot | Exceptions, rule fixes, sample return | Replay localization, metrics, failure feedback |
| Release | Change flow, model or roles | Revision bound to Run, canary, audit, rollback |

## For business teams

- Replay by business order: context, tool sequence and versions are visible
- Outbound and writes freeze before approval—no silent model sends
- Gray and high-risk steps carry human gates; rules inherit across scenes
- New scenes reuse proven failure handling instead of rediscovering edges
- Connect drawing / vision / X-ray models through one tool surface

## For managers

- Deliverables fit ops manuals and audit trails: versions, gates, freeze snapshots, rollback paths
- Turn “uncontrollable AI” into signable acceptance items
- Observe success rate, human-intervention rate and outbound blocks
- One foundation compounds across product lines—no chimney rebuilds
- Control and customer acceptance share one ownership chain

## Typical scenarios

| Scenario | Where the non-substitutable edge shows |
|---|---|
| Cross-system workflows | Mail, forms, approvals, writeback; freeze + idempotency; compensate and replay by order |
| Station QC / film reading | Schedule vision or X-ray models; gray-zone review and dual-sign; evidence to QMS/MES |
| Drawing to process / BOM | Chain drawing models and rules; escalate low confidence; field–geometry lineage |
| Robot task loops | Step perception–decision–execution; force gates and human takeover on hazardous acts |
| Multi-role delivery | Manager / worker / auditor; Gate accept/reject drives the remediation loop |

## Capability snapshot

| Capability | What it does |
|---|---|
| Harness loop | Goal → plan → tools → observe → finish or escalate |
| Gate & freeze | Snapshot first, approve, then execute high-risk actions |
| Order-level replay | Tool sequences, HITL decisions and outbound indexed by order ID |
| Multi-agent | Roles, delegation, parallel work and merge |
| Tiered memory | Session vs long-term; searchable and expirable |
| Sandbox | Allow-lists for files, network and credentials |
| Skills / Tools | Shared I/O and permissions against systems and domain models |
| Governance | Versioning, audit, observability, cost control and rollback |