Hard budget actions
Caps can alert, move work to a lower-cost model, or pause execution instead of merely reporting spend after it happens.
How Grex works
Trust the outcome, not the agent.
AI agents can misunderstand instructions, repeat work, overspend, or declare victory too early. Grex is designed around that reality: consequential work gets boundaries, evidence, independent checks, and an operator-visible decision about what actually happened.
Designed for imperfect agents
Grex assumes agents will misbehave.
The goal is not to pretend every model will behave perfectly. The goal is to keep a bad turn, a weak result, or an endless retry from becoming unlimited spend or accepted success. Workers can propose and produce. The control plane decides whether the work satisfies the mission contract.
The mission lifecycle
From an objective to a checkable outcome.
- 01
Contract the outcome
Define what success means, how long the mission may run, what it may spend, and when it should stop.
- 02
Dispatch bounded work
Grex assigns work to capable agents across the computers you control without making one chat session the system of record.
- 03
Persist the evidence
Artifacts, costs, claims, checks, and material state changes remain attached to the mission as durable records.
- 04
Challenge the result
Verification evaluates the evidence against the contract. A confident sentence from the producing agent is not sufficient.
- 05
Ratify or stop
The mission produces a ratified outcome, waits at an operator gate, or stops with evidence explaining why it should not continue.
- 06
Issue the receipt
The terminal outcome becomes a portable, signed record that can be checked without trusting a chat transcript.
When an agent goes wrong
Failure becomes bounded and visible.
| Agent behavior | Without mission controls | With Grex |
|---|---|---|
| Keeps trying | More calls, retries, and spend | Spend and mission boundaries limit the blast radius |
| Claims success too early | The answer sounds complete | Missing evidence prevents ratification |
| Loses context | The conversation becomes the recovery plan | Mission state, artifacts, and receipts remain durable |
| Needs a consequential decision | The agent guesses or waits in an ephemeral session | A durable operator gate holds the decision |
| Cannot produce enough value | Activity may still look like progress | The mission can end with an evidence-backed stop |
Controls that survive the agent
The safety model lives outside the prompt.
Mission economics
Time, spend, minimum value, margin, and operator-attention terms make business usefulness part of the definition of done.
Independent checks
Verification and prosecution challenge outcome claims against durable evidence before those claims can become authoritative.
Operator gates
Policies can reserve consequential judgments for the operator while unrelated work continues elsewhere in the fleet.
Away-readiness
Grex checks for an enforced budget, reachable notifications, a live heartbeat, and approved work intent before calling a node ready.
Verifiable receipts
Terminal mission receipts and ratified artifacts can be exported as signed documents and verified offline.
The result
“Done” is a claim. A Grex outcome is a claim with evidence, judgment, and a receipt.
Grex does not promise perfect agents. It prevents imperfect agents from defining success, spending without limits, or failing without leaving an inspectable record.
Frequently asked questions
How Grex controls autonomous work
Does Grex prevent every agent failure?
No. Grex is designed to bound failures, preserve their evidence, and prevent an unsupported result from becoming an accepted outcome.
How does Grex control runaway spending?
Grex supports enforced budget caps that can pause work or move it to a lower-cost model. Missions also carry their own spend ceilings.
How does Grex decide that a mission is complete?
Required artifacts and checks are evaluated against the mission contract. The result is ratified, held for an operator decision, or settled with evidence explaining why it stopped.
Can I use my current AI assistant with Grex?
That is the product direction. Grex is moving toward assistant-agnostic control over MCP so assistants such as OpenClaw, Hermes Agent, and Grok Bot can become command surfaces rather than replacements for Grex.
What consequential outcome would you delegate if the work came back with evidence?
Share your interest