A governed foundation for AI agents.
Let AI agents do real work in your business, within the rights you grant, in the order you allow, with a signed record of every move.
The rulebook, the order of moves, the undo, and the history are already built. The agent itself, and the controls shaped by how your teams work, are built with you. This page says which is which.
An agent proposes. The platform decides what is allowed, and remembers what happened.
Every workflow follows your rules
You define the moves a workflow can make and who may make each one. Anything else is refused, and every move is kept in a history that is only ever added to.
Agents get exactly the rights you grant
Each agent has an account of its own, governed by the same groups, roles, and permissions as your staff, down to individual fields. It outlives any one employee, is reported on separately, and can be switched off at once.
Multi-step actions undo themselves
When an action touches several systems and a later step fails, the completed steps are reversed. Actions are safe to retry, so a half-finished job never lingers.
Work carries on in the background
Delayed and long-running steps are queued, retried on failure, and set aside for review if they keep failing, rather than lost.
Supplier invoice reconciliation
Started by a new invoice · agent: Invoice Assistant
Queued: invoice INV-7781 received
Planned: three matching steps proposed
Approved by A. Reviewer
Step 1: matched to purchase order
Step 2 refused by the ledger; step 1 reversed
Five stages. Three run on the platform today; two are built for you.
The order is the point: nothing changes before it is approved, and nothing is lost if a step fails.
Triggered
A record change or a timer starts the run
Planned
The agent proposes the steps
Approved
A person signs off before anything changes
Executed
Steps run; a failure reverses the rest
Recorded
Every move kept, with who made it
Most of it is already there. The agent layer is the build.
Configuration is our work but not new software: it is set up for you and changed through a planned update, without new code.
In the platform
Already built
- Accounts for agents, governed like staff accounts
- A standard connection that assistants such as Claude and ChatGPT use to read and act on your records
- The same permission checks for agents as for people, down to single fields
- Workflows that move only in the order you define, with their full history
- Multi-step actions that reverse themselves when a step fails
- Background work with automatic retries
- Agent output searchable and reportable within a second
- Cited answers from your own documents
Configured for you
Set up, not written
- The records your workflows need: runs, steps, and sign-offs
- The allowed moves of each workflow, and who may make them
- Which actions each application offers to agents
- Which documents an assistant may draw on
- The groups, roles, and permissions each agent account holds
- Spending limits per account and per organization for the assistant
Built for you
New work, on top
- The agent itself, or the connection to the agent you already have
- An approval hold that parks an agent's action until a person signs off
- A step-by-step view of each run, with a spending limit per run
- The specific actions your workflows take in other systems
- Deadlines and escalation for steps left waiting
- Search by meaning over your documents, offered to agents as a tool
The controls you would otherwise build around an agent.
Permissions
One rulebook for all
Actions an account may not take are hidden or refused, for agents and people alike, and each one is checked again as it runs.
Isolation
Your own partition
Every record carries its organization, and the platform will not touch one without naming it.
Untrusted input
Anonymous input labeled
Public submissions reach an agent marked as untrusted, so it can treat them as material to read, not instructions to follow.
Activity record
Every agent action signed
Each agent action is logged against the agent's own account, and the log is signed hourly, so any later tampering shows.
Recovery
A wrong run is a rewind
Data can be restored to any point in the last 35 days, so an agent that changed the wrong records is a recovery, not a loss.
Grounded answers
Cites, or declines
The built-in assistant answers only from documents the account may see, names its source, and declines rather than guesses.
See it live in AtlasWorth knowing before you budget.
The agent is built for your workflows. The loop that decides what an agent does is built with you, or brought from the agent framework you already use.
Approval rules are yours. The hold that parks an agent's action until someone signs off is built to match who approves what in your organization.
Costs are set per run. A step-by-step view of each run, and the spending limit it may not exceed, are built alongside the agent.
Deadlines follow your service levels. Escalation for steps left waiting is set up to the timings your teams already work to.
Decide versioning at design time. Agree up front how a change to a workflow's rules should treat runs already in progress.
Hidden instructions are a shared job. The platform labels anonymous content; screening what agents read is part of the agent's design.
Coda
Walk one workflow through.
If your agents need to change real business records under real permissions, with a trail you can stand behind, the hard parts already exist.
What you would budget for is the agent layer on top. Bring us one workflow: walking it through the five stages takes about a day, and shows exactly which pieces it needs.