A forward-deployed engineer for AI agents comes into your company and builds agents that take real actions there: cleaning up cloud resources, changing infrastructure and handling back-office work. That is how we use the term for our own engineers, and on our engagements every action needs independent sign-off and leaves a receipt anyone can check. The work is mapping the permissions each agent needs, building a gate with an author, a reviewer on a different model and a person's passkey for production, testing the reviewers by planting mistakes, and handing over a runbook and an evidence pack mapped to the controls your auditor tests.
When should you bring in a forward-deployed engineer?
We see two problems stop teams that want agents to act. First, most "multi-agent review" is one process wearing several hats. When an auditor or a post-incident review asks who independently approved a change, there's no answer you can prove.
Second, nobody inside has done it before. Wiring agents into real systems with the right permissions, reviewers and evidence takes engineers who have run agents in production. If your team has the first problem and not the second, you may not need outside help; the checks below are a good place to start either way.
What does a forward-deployed engineer do, stage by stage?
We run a gated agent deployment in four stages over eight weeks.
- Weeks 1 and 2, pick the jobs. Choose one or two jobs worth automating, and map the permissions each agent needs, and nothing more. You approve those permissions before anything is built.
- Weeks 3 to 5, build and gate. Build the agents and set the rules: an author, a reviewer on a different model, and a person's passkey for production.
- Week 6, test the reviewers. Plant mistakes and check the reviewers catch them, so approvals mean something.
- Weeks 7 and 8, go live. Hand over the runbook and an evidence pack mapped to the controls your auditor tests.
How do you test a reviewer's ability to catch mistakes?
You can run this test this week. Take a change an agent would normally propose, like a Terraform plan or a rightsizing recommendation, and insert a deliberate error: a missing resource, an incorrect region, or a permission that exceeds scope. Do it on a throwaway branch against a sandbox account, with apply and merge turned off for that branch.
Run the agent's review process against it. If the reviewer doesn't flag the error, the gate isn't working. Repeat with different errors until the reviewer flags them consistently.
The planted change never merges or applies, even after the reviewer catches it. If you plant the error by deleting a resource from the configuration, the plan shows it as a destroy, and an apply would remove the real resource. When the test is done, close the branch or pull request and check your pipeline logs to confirm no apply ran.
What permissions should an agent get?
Our rule is read-only first, scoped per agent after that. Permissions are designed early, and the client approves them before anything is built.
To check your own agents, list every permission each one holds, whether it runs now or is planned. Mark the ones its job needs. Everything unmarked is risk that buys you nothing.
Be careful tightening scopes on a live agent. If you remove a permission it still needs, its next run may start a change and fail halfway through, leaving partial state. Test permission changes in a staging environment first. Before each cut, export the agent's current policy to a file and confirm you can restore it from that file. Keep that copy until the agent has run cleanly on the new scope.
What jobs come first, and what does the engagement include?
Typical first agents are cloud cleanup and rightsizing, Terraform plan and apply, infrastructure pull requests, and back-office steps with a maker-checker rule. We're model-neutral: the reviewer is deliberately on a different model from the author, and we plug in the tools you already use.
Our engagement is fixed scope and fixed price. It includes one or two agents in production that take real actions, approval rules with an independent reviewer and a human passkey for production, a reviewer red-team with planted mistakes, SOC 2 change-management evidence for every gated action, and a runbook and handover to your team. It is built on Seconded, our platform for separation of duties between AI agents, which stays after we leave so the controls keep working.
Tell us the job you'd hand to an agent if you could trust it, and we'll tell you how we'd gate it and whether eight weeks is realistic: book a 30-minute call.
Frequently asked questions
What is a forward-deployed engineer?
No primary source defines the term. We use it for an engineer who works inside a client's company, on its systems, building AI agents that take real actions, where every action needs independent sign-off and leaves a receipt anyone can check.
Do I need to use a specific AI model?
No. We're model-neutral. The reviewer is deliberately on a different model from the author, and we plug in the tools you already use.
How is this different from a CI/CD pipeline?
A pipeline runs your changes. The gate decides whether an agent's change may run at all: an author, a reviewer on a different model, and a person's passkey for production, with change-management evidence for every gated action, so you can show who independently approved it. In practice the gate sits in front of the pipeline you already have.
How long does it take to get AI agents into production?
Our engagement runs eight weeks, from picking the jobs to agents in production.
This isn't compliance advice.