OSuite · Agent Action Control Pilot
Your agents can act. Can you control what they do?
A focused, 14-day paid pilot that adds approval checks, execution controls and audit evidence to one AI agent workflow, without rebuilding the agent.
One workflow. One action boundary. Clear acceptance criteria.
Example
A refund is approved. Then the agent changes it.
-
01 · Agent
Requests a US$5,000 refund
refund.issue · $5,000 · ca_7b2c…0af
Requested -
02 · OSuite
Requires approval
Over the US$1,000 auto-approve limit. Nothing runs yet.
Waiting -
03 · Manager
Approves US$5,000
Approval bound to ca_7b2c…0af
Approved -
04 · Agent
Changes the amount to US$8,000
refund.issue · $8,000 · ca_e41d…9c3
Changed -
05 · OSuite
Blocks execution
ca_e41d…9c3 ≠ ca_7b2c…0af. Nothing runs, and the attempt is recorded.
Blocked
Approval for one action is not permission for another.
A pause-for-approval step stops a workflow until someone clicks approve. OSuite binds the approval to the exact action, so if the amount, the destination or anything else changes after approval, the approval no longer applies. A simplified illustration of what the pilot tests.
The problem
The problem starts when agents stop answering and start acting.
Even a correct agent action may need authorization. Understanding a request is not the same as being allowed to carry it out.
Agents that refund payments, cancel orders or change customer details.
Add approval to the actions that need it.
Agents that change production configuration or run commands.
Put an execution boundary around high-risk actions.
Agents that update CRM, ERP or other business records.
Keep a record of every governed action that can be checked later.
What you get
A production-oriented control layer for one agent workflow.
-
01
Action boundary
We identify and connect one execution boundary to govern, such as an agent calling a refund API or triggering a production deployment.
-
02
Policy checks
Rules that allow, deny or require approval for the actions in scope. For example: small refunds run automatically, large refunds need a person, unauthorized refunds are refused.
-
03
Human approval
An approval path for the actions that need a person. The approval binds to the action that actually runs: approving one action never authorizes a different one.
-
04
Execution evidence
A record for every governed action: who started it, what was requested, which policy applied, whether it was approved and what happened. An outcome that is unknown stays unknown; it is not recorded as success or failure.
-
05
Validation report
The policies tested, refused and approved cases, execution records, open issues and recommendations for production. Something you can take to your CTO, your CISO or your own client.
What you’ll use
OSuite as it ships today, at one action boundary.
No new agent framework and no rebuild. OSuite connects where your agent already acts.
Connector
The installer, the osuite CLI and SDK, an MCP or API gateway route, or the n8n and Dify templates: whichever surface your agent already uses.
Policies
Rules for the actions in scope, with routing signals such as blast radius, destination and reversibility.
Studio
Approvers see the exact action, its risk and its authority before it runs, then replay any decision afterwards.
Receipts
Every decision closes into a tamper-evident receipt bound to the action, exportable as a signed packet for an auditor or a client.
Is it a fit?
Built for one real workflow, not a platform rollout.
A good fit
- An agent that takes actions in an external system, such as refunds, deployments or record updates, or one that is blocked from doing so.
- One workflow you can name, and one owner who can agree the scope.
- A test environment, a sandbox or scoped credentials we can work with.
- You need approval, refusal or audit evidence before going further.
Not a fit
- Agents that only answer questions or draft text.
- Governing every agent in the company at once.
- A security audit or a compliance certification.
- Guaranteed outcomes in systems neither of us can observe.
How it runs
Fourteen days, fixed scope.
One workflow and one main action boundary, so the pilot stays a pilot. The 14 days start once the scope is agreed and the access we need is in place.
Scope
Choose one workflow and its key action.
Pilot scope
Integration
Connect OSuite at the execution boundary.
Working integration
Policy
Configure the approve, deny and allow rules.
Policy configuration
Validation
Run the test cases.
Test evidence
Review
Demo, retrospective and handover.
Validation report
Acceptance criteria
What we test against.
These are the targets we agree with you during scoping, not guarantees for every system. Each one is confirmed against your integration before it goes into the pilot agreement. Anything we cannot deliver reliably in your environment stays out of it.
- Authorized action
- Runs.
- Unauthorized action
- Is refused.
- Action that needs approval
- Does not run before it is approved.
- Action changed after approval
- The original approval does not authorize the changed action.
- Execution times out
- The outcome stays unknown; it is not reported as success.
- Audit query
- Key decisions and execution records can be traced.
Price
US$2,500
Fixed-scope pilot · 14 days · 1 workflow · 1 action boundary
50% when we sign, 50% on acceptance. Before anything is signed, a short technical feasibility check confirms the pilot can be delivered in your environment.
Afterwards you choose: keep the governed scope running on a monthly plan, or extend it to more agents, action boundaries, policies and environments. We agree pricing for either once the pilot is done.
Why a fixed pilot
- No platform contract, seats or minimum term before you see results.
- Half the fee is due only on acceptance, against criteria agreed up front.
- You keep the validation report and test evidence whatever you decide next.
- Approval binding and replayable evidence are a project of their own to build in-house. The pilot puts both in place for one workflow and shows whether they are worth extending.
Included
- One technical scoping assessment
- One execution boundary integration
- Up to three core policies
- One human approval path
- One set of acceptance tests
- One validation report
Not included
- Building a multi-tenant enterprise platform
- Company-wide agent governance
- Custom compliance certification
- Unlimited API integrations
- Round-the-clock operations
- Guarantees against all business losses
Why trust it
Safe to try, built on research, already in use.
Safe to try
- Least privilege: a test environment, or credentials that stay under your control.
- High-risk actions wait for a named person before they run.
- SSO, SCIM and role-based access connect to your identity provider.
- In production, OSuite can run hybrid or air-gapped, inside your own cloud, VPC or on-prem.
- Security package, subprocessor list and DPA on request, under NDA.
Built on research
- Approval binding comes from our work on PCAA, and canonical action identity from CAVA.
- OSuite is the reference implementation of Ambit, an open protocol. Its published layer is MIT licensed and testable without an OSuite account.
- CAVA and PCAA are cited in two independent research preprints.
- Our papers are arXiv preprints, not peer reviewed.
Already in use
- Live in Claude’s connector directory, as a ChatGPT app, as an n8n template and in the Dify marketplace.
- 10K+ downloads.
- An enterprise proof of concept under way with Mercari.
- ACSC Network Partner in Australia. Backed by NVIDIA Inception and Block71.
- Built by Ond, a team of researchers and engineers. The founder previously worked at Apple and Microsoft and helped take Dify Japan from zero to $1M ARR.
Questions
Before you apply.
Do I need to replace my agent framework?
That is the goal. It depends on whether your workflow has an execution boundary OSuite can connect to, which we check with you before anything is signed.
Is this a security audit?
No. It is an execution-governance pilot for one workflow. It does not provide a full security audit or a compliance certification.
Do you need access to production credentials?
We prefer least privilege: a test environment, or credentials that stay under your control.
Where does OSuite run during the pilot?
Usually as the hosted Studio with a connector in your environment. If your data cannot leave your infrastructure, tell us at scoping: a hybrid or air-gapped setup changes the scope.
What do you need from our team?
Someone who knows the workflow for scoping and the final review, and access to a test environment or scoped credentials for the integration.
What happens after 14 days?
You get the validation results and decide whether to keep the governed scope running, extend it, or stop there.
Apply
Apply for a pilot.
Tell us about one workflow. We review every application and reply within two business days.
Prefer email? access@osuite.ai
Thanks. We’ll review your workflow and respond within two business days.
Reference
OSuite is the reference implementation of Ambit, an open protocol for agent actions. Its first layer, Action Effect Spec, is a published draft; the other layers are research that already runs in OSuite. The pilot governs one workflow with OSuite; it does not deliver a complete Ambit implementation.
Our research and Action Effect Bench results are preprints and benchmark runs, not customer deployments.