AI Agents on Government Programs

Government program teams are being asked for two things at once: deliver faster, and be able to show exactly what changed, who approved it and why. AI coding agents look like they help with the first and put the second at risk. An agent can produce a working change in minutes. It can also produce a plausible change nobody fully understands, reach for files and tools it shouldn’t touch, or send code and prompts somewhere they were never allowed to go.

The answer isn’t to keep agents off government programs. It’s to treat an agent as one more contributor inside an engineering process that was already built for accountability. At Thazen, we’ve found that the value of agentic AI comes less from the model than from the loop around it.

Start With the Loop, Not the Model

At Thazen, every piece of work runs through the same four steps, whether an engineer, an AI agent or both are doing it. It’s the same method our engineers use on Army training software and on AI systems alike.

Plan. Scope the work and write down the approach. The plan is read-only until a person approves it.
Act. Make the change in an isolated workspace, with scoped permissions and allow-listed commands for any agent.
Observe. Run the tests, read the logs and check the results, inside the program’s environment.
Evaluate. Score the result against tests and evaluation harnesses, then a person reviews the diff.

None of this is new. It’s how good engineering teams have always worked. What’s new is that an agent can go around the loop very quickly, which makes the checkpoints more important, not less.

Every Change Is Reviewable

An agent’s work should arrive the same way a colleague’s does: as a diff or pull request that a person reads and approves. Nothing merges on its own. That’s what we mean by work you can review.

That sounds obvious, but it shapes how you set up the work. Keep changes small enough that a reviewer can actually read them. If an agent hands back a change too large to review with care, the task was scoped too broadly, and the fix is to split the task, not to skim the diff. The review record then does what program oversight needs: it shows who looked at the change and who accepted it.

Tests Before Trust

“The tests pass” isn’t enough evidence on its own, especially when the same agent wrote the code and the tests. Our rule is simple: a change isn’t done until a test proves it, and that test fails without the fix. A test that fails first and passes after gives the reviewer real evidence rather than a green checkmark.

Pipelines run those tests, plus security scans, on every change. On legacy systems, where test coverage is often thin, the first job is usually to write tests that capture what the system does today, so that every later change, human or AI, has something to be checked against. Evaluation works the same way for AI itself: on I2ST, a GenAI training prototype led by Design Interactive, our engineers scored each AI conversation against a rubric, so the output became structured, reviewable feedback rather than an opaque score.

Guardrails Set Up Front

The time to decide what an agent may do is before it runs, not after something surprising shows up in a log. That means tool permissions, allow-listed commands, a workspace it can’t escape, spending limits on model use, and an audit trail of every action it took.

Federal and DoW programs already think this way about people and systems: least privilege, separation of duties, a record of who did what. Agents should sit inside those same rules, under the program’s own security and delivery policies, rather than getting a special exemption because they’re new.

Your Boundary, Your Data

On a government program, the question of where code, prompts and outputs go isn’t a detail. It decides whether you can use the tool at all. Models should run in environments the program controls, such as Amazon Bedrock in AWS GovCloud (US), so that nothing leaves the boundary on its way to a model. That requirement is why our own engineers work with Topsoil, the private coding agent we built to run in the customer’s own AWS account. Backlog items come back as pull requests for a person to review, and Topsoil never merges on its own. The principle matters more than any particular tool.

The same goes for AI that answers questions rather than writes code. Answers should cite their sources, so a person can check the claim against the document rather than trusting the summary. On work for the U.S. Courts – Ninth Circuit, that meant machine learning that helps analysts keep up with content volumes, with people still making the call.

Engineers Stay Accountable

AI changes how an engineer spends the day: less time typing, more time planning, reviewing and testing. It doesn’t change who answers for the result. When a change ships, a named engineer approved it and can explain it.

That’s also why seniority matters more with agents, not less. Judging whether a confident, well-formatted change is actually right takes experience with the system, the domain and the ways software fails. At Thazen, we’d rather measure an agent-assisted team by what it delivers and how cleanly it can show its work than by how much code it generates.

Where to Start

You don’t need a program-wide rollout to learn whether this works for you. Pick one backlog, one legacy component or one well-bounded AI idea. Agree up front on the gates: what the plan must contain, which tests must exist, what the agent is allowed to touch and who reviews. Start with plans read-only, keep the changes small, and look at the review record after a few weeks. It will tell you more than any vendor demo.

How Thazen Can Help

AI-accelerated. Engineer-led. That’s the balance we think government programs need: use AI to move faster, and keep engineers accountable for everything that ships. Thazen’s senior engineers work embedded in program teams, often as a subcontractor alongside the prime, and bring the Plan, Act, Observe, Evaluate loop with them. If you’re working out how to bring AI agents into a federal or DoW program without losing control of it, we’d be glad to compare notes and scope a first loop with you.

AI Agents on Government Programs: Faster, Without Giving Up Control