Every control, enforced and checkable
For security and risk teams: how what your company decides is enforced on every run. The plan signature, the pre-run check, the run record as evidence and what stays protected while the agent works.
The same rules on every run
Every task goes through the same controls, limits the agent cannot skip: permissions, a check before it starts, a Sealed Plan, approvals and a run record at the end.
The controls on every run
- 01
Permissions
What an admin turns on
- 02
Pre-run check
Agent, permissions, accounts and where it works
- 03
Sealed Plan
If altered, it won’t run
- 04
Human approval
Exact values, by whoever you choose
- 05
Execution
In an isolated folder, only with the Sealed Plan
- 06
Run record
Who, what and with which permissions
Your team sets permissions, Two-Key Approval, per-system limits and stopping an agent in the platform. See what you decide
Checked before it starts
Before a run starts, the platform checks that the agent is active, that its permissions cover every step and that the operator’s accounts and the place where the agent will work are ready. If anything is missing, the run doesn’t start.
- The platform signs every Sealed Plan with Ed25519, and the Reelease Agent verifies the signature, fingerprint and expiry before running it. If anything doesn’t match, it doesn’t run
- The signing key sits in a secret store managed by Reelease; agents and operators never hold it
- During the run, every action on your systems and tools is checked again against that same plan
- Using the operator’s computer always needs their consent, and you can require it for the browser and web forms too
The run record as evidence
Every run leaves its run record: who started it, the plan, each approval with who and when, every action with its result and the version of the rules that applied.
- SHA-256 fingerprints of the plan, of each approved change and of the signed manifest, which you can check in your browser
- It exports and is kept even if you delete the agent, with retention periods per system and legal hold
- The sample run record is public: you check its fingerprints and its manifest signature without installing anything
- Lucía Prieto launches the task
- Lucía Prieto reviews and approves the plan
- Andrés Molina approves change RTDC
- Andrés Molina approves change RTRY
- Run completed
Protected while it works
- Where the agent works
- The agent works in an isolated folder for each run, in the cloud or on the operator’s computer. It only runs Sealed Plans, signed by the platform, and holds no credentials or AI model keys.
- Credentials
- Each operator’s access and each integration’s keys are stored encrypted (AES-256) and never reach the operator’s computer. Neither the agent nor the model sees them.
- Access and identity
- Your team signs in with your identity provider (Pilot, Pro and Enterprise) and with a second factor on every plan. Towards your systems, a technical credential per connection is used by default; who asked and who approved travel with every call and stay in the run record.
Frequently asked questions
What is the Sealed Plan?
The platform builds the plan from your published executable use cases and your rules, starting from what the model proposes. It seals it with a fingerprint, signs it and signs each step’s data too. That’s how it reaches wherever the agent works, in the cloud or on the operator’s computer. If it’s altered or arrives too late, it doesn’t run, and every action is checked against that same plan.
Who signs the plan?
Reelease signs every Sealed Plan with Ed25519. The key sits in a secret store managed by Reelease, agents and operators never hold it, and it rotates with an overlap period. The Reelease Agent rejects a plan whose signature, fingerprint or expiry doesn’t match.
What is checked before every run?
That the agent is active, that its permissions cover every step within its limits, that the place where it will work (the cloud or the operator’s computer) is ready, and that the operator’s accounts have the permissions the task needs. If anything fails, the run doesn’t start and the operator sees what’s missing.
What is recorded for each run?
Who started it, every step, every approval (who gave it and with what comment) and every action on your systems, with the version of the permissions that applied. It can be exported as a run record and is kept even if the agent is deleted, with retention per system and legal hold.
And configuration changes?
They go into the activity log: permission and configuration changes, agent templates, integrations, data retention, Bridge Admin actions and every approval, including rejected, expired and emergency ones. You can filter by type.
What stops a prompt injection?
Data from your systems and documents reaches the model marked as data. The real barrier is structural: the model cannot add anything unpublished to the sealed plan or skip an approval.
Can my team sign in with our company’s identity provider?
Yes, on Pilot, Pro and Enterprise: Microsoft Entra ID, Google Workspace, Okta or any OpenID Connect provider. You verify your domain and your team signs in only through your provider, under your own access policies; only accounts from that domain get in, and when someone leaves your company, their access closes within 24 hours at most. The second factor is on every plan and you can require it for your whole team.