Skip to content

You decide what every agent does

Every task follows a Sealed Plan: if the plan or its values change, it doesn’t run. With Two-Key Approval, every change waits by default for someone else, who approves it with the exact values in front of them. And every run leaves its run record.

We email you the link to open on your computer.

Where the model stops

The language model helps understand what you ask and proposes the steps. What reaches your systems is decided by the Sealed Plan, the approved values and your rules.

  • Model
  • Reelease Control
  • Your system
  1. Model: Proposes

    Reads the request and suggests the steps. With a published executable use case, the plan comes ready-made.

  2. The model stops here
  3. Reelease Control: Builds and checks

    Only with published capabilities. Each step declares whether it reads or changes, and your rules check it first.

  4. Reelease Control: Seals

    The plan and its values are fixed and signed. Changes wait for Two-Key Approval.

  5. Your system: Receives

    A call pinned to the sealed step, with the exact approved values for changes.

What each agent can do, one capability at a time

A new agent can’t touch any system. An admin turns on, one by one, the capabilities it needs: your systems, email, documents, the browser or the operator’s computer, the last one always with their consent.

  • It only calls what your team published as an executable use case
  • What the system’s rules forbid, no agent can allow
  • Every plan has a cap on steps and changes: 30 and 10 by default
  • Every rule change is recorded with who, when and why, and you can go back to an earlier version

Two-Key Approval, with the exact values in front of you

By default, anything that changes data waits for someone other than the person who started the task, who approves it with the exact values in front of them. The approval is bound to those values: if they change, it no longer counts. An admin can relax it for an agent, and the reason goes into the run record.

How a change goes through

  1. 01

    Start

    A person starts the task. With Two-Key Approval, by default, they don’t approve their own changes.

  2. 02

    Your rules

    Before each change, your rules check it and decide who approves it.

    Above your cap: If the amount exceeds the system’s cap, two people approve it or it’s rejected, as you configure.

  3. 03

    Another person approves

    They see the exact values of the change. If those change, their approval no longer counts.

  4. 04

    Notice and deadline

    The notice lands in their inbox and, if they’re not on the platform, by email too. Before the deadline, a reminder.

    If no one answers: When the deadline passes, the approval expires and the change doesn’t go out.

  5. 05

    Run record

    Who started it, which rule applied, who approved and the fingerprint of the values.

Exact values

Whoever approves sees the exact values, and only those run.

See it as:
Approval required
Record the source-of-funds proofChange
Approval for this changeYou approve only this change, with these exact parameters. If they change, a new approval is requested.
Action
RTDC · Core banking, CICS
Parameters
transferIdTRX-RET-9002
documentTypeSOURCE_OF_FUNDS
documentRefDOC-88412
Parameters fingerprint
e23737b7f0fe21a2aac075a49d9e3bb1…
Lucía PrietoStarted the taskAndrés MolinaAssigned · Operations leadAwaiting your decision
Cancels whatever is running

If the values change, a new approval is requested. The record keeps their fingerprint, without the data.

See this approval in the record

Two-Key Approval

The rule every system applies.

It’s the default rule, versioned on each system. If an admin relaxes it for an agent, the reason goes into the run record.

Owner and deadline

Once the deadline passes, it can't be approved.

Each step can have an owner and a deadline. The alert lands in the operator's inbox.

Information mid-task

If information is missing, it asks you.

Your answer is checked against the fields the step declares before it carries on.

Emergency list

It resolves other people's tasks, never its own.

Always with a reason and their name in the record. Rejecting asks for a reason too, by default.

Email alert

They get an email and, before the deadline, a reminder.

It's sent when the approver isn't on the platform. The email never carries the values.

Activity panel

You see how long approvals take.

Per agent, the pending ones and the average time. Per system, the rejected, the expired and those resolved through the emergency list.

Run record

Every decision, with its rule and its reason.

  1. Lucía Prieto approved the plan“Plan reviewed: two reads and two changes, RTRY last.” · plan fingerprint 73f640c5ed65…
  2. RTDC is waiting for Andrés Molina

Who approved, under which rule and why, and the fingerprint of the values without the data. It can be exported and is kept even if you delete the agent.

Open it

Limits per system

For each system you set change windows, freezes and business limits. Lookups keep working; changes go out only within what you set.

  • Change windows: the days and hours when the system can be changed, in its own time zone
  • Dated freezes, such as a month-end close, when no change goes out
  • Outside the window, the change doesn’t go out: the task stops or waits for a person to send it once the window opens, as you choose
  • Business limits per change (amount and number of records) and per day (number of changes)
  • If a change goes over a limit, it is rejected or needs approval from two different people, as you configure it

The run record of a task

Every run leaves one: the checked plan, who approved what and when, every action on your systems and its outcome. This one comes from a sample banking run: pick a step.

Record the source-of-funds proof that has already arrived for TRX-RET-9002 and release the transfer.

Banking core in COBOL and CICS at Bank of Z, a sample bank

Completed
Agent
Banking operations
Started by
Lucía Prieto · Back office
Approved the changes
Andrés Molina · Operations lead
Duration
6 min 3 s

Plan ·

Plan proposed, checked and signed

The agent proposes 2 reads and 2 changes. Reelease checks the agent and its rules and signs the plan: if anyone alters it, it doesn't run. Whoever starts the task reviews and approves it.

Proposed by
The agent, with the system's AI model
Steps
4 (2 reads and 2 changes)
Plan approved by
Lucía Prieto · 10:42:31
Plan hash
73f640c5ed65…

A sample banking task, with demo data and no personal data. Every task your agents carry out leaves a run record just like this.

What happens when you suspend an agent

Any member with full access to the platform and the agent can suspend it. Whatever is running is canceled, the agent can no longer reach your systems and nothing new starts.

  • Only an admin can resume it
  • You can also cancel a single run
  • If a step fails after a change and the contract declares how to undo that change, Reelease proposes the compensation with exact values and a person approves it

What an agent cannot do

The platform checks it on every task, whatever the request says.

  • Call what you didn’t publish

    The step is refused before it reaches your system.

  • Touch a system you didn’t turn on

    If an admin didn’t turn that capability on, the pre-run check won’t let the task start.

  • Approve its own change

    By default, a change waits for someone other than the person who started the task.

  • Send a value other than the sealed one

    If a value changes after approval, the approval no longer counts and a new one is requested.

  • Work without a run record

    Every run leaves its own: the plan, each approval and each action with its result.

  • Keep going once suspended

    Whatever is running is canceled and nothing new starts.

For your security and risk team: how each limit is enforced, with the plan signature, the pre-run check and the run record as evidence.

See how each control is enforced

Frequently asked questions

Can a new agent act as soon as it’s created?

No. It starts with everything off (least privilege) and can’t touch any system. Only an admin turns on capabilities, by hand or with an agent template.

Who approves what the agent does?

Whoever you choose, agent by agent: which steps wait for a person, who approves, by when and what is pre-authorized. By default someone else approves, not whoever started the task, and rejecting requires a reason. If an admin allows it, whoever starts the task can approve it by giving a reason, which goes into the run record. Approvers see the exact values, and only those values run.

Can an agent get around a system’s limits?

No. Change windows, freezes and limits are checked before each real change is sent to your system. If the system’s rule can’t be checked, the change doesn’t go out.

What if a step fails halfway through a task?

Every change carries an idempotency key. If a call times out, the step is marked as outcome unknown and can be checked before retrying. If the contract declares how to undo that change, Reelease proposes the compensation with exact values and a person approves it.

Can an agent be stopped immediately?

Yes. Any member with full access to the platform and the agent can suspend it: whatever is running is canceled, the agent can no longer reach your systems and nothing new starts. Only an admin can resume it. You can also cancel a single run.

See it on your own system

We’ll show you Reelease on one of your cases: what it understands, the plan an agent would propose and what you’d sign off on.

You choose the cookies

We use technical storage to remember your choice and, only if you accept, analytics cookies (Google Analytics) to understand how the site is used. You can change your choice at any time. Learn more in our cookie policy.