AI Coding Agents

Five Ways to Make AI Coding Agents Easier to Trust

When an AI coding agent says a feature is done, how much work does it take to find out whether it is? In Matt Krainski's Ottawa talk, the best answer is a short recording attached to the pull request. A reviewer can see the behavior before opening the diff, then read the tests and inspect the code with a clearer idea of what to look for.

That small habit captures the point of his five practices: give the agent enough room to build and test, but make its actions easy to contain and its results easy to review. The setup matters as much as the instructions you give Claude Code or Codex.

Watch the Ottawa Talk

Source: Power Up Your AI Coding Setup by Matt Krainski of Gravital Forge, recorded at AI Agents Ottawa #3 on November 18, 2026. The checklist and copyable prompt below are JQ AI SYSTEMS adaptations of the talk.

1. Give Agents a Dedicated Workspace

Krainski prefers a separate machine or cloud VM for longer sessions. The machine can keep running when he closes his laptop, and the agent has a predictable place to build the app, run tests and start a test database. The important part is what the environment can reach, not where it sits.

Start with a non-admin account, a clean checkout, test data and project-scoped credentials. Keep production secrets, personal browser sessions and unrelated projects out of reach. A Git worktree helps two sessions avoid editing the same checkout; it doesn't isolate their permissions. Keep tool-level sandboxing and approvals as well. The Claude Code sandboxing guide and OpenAI's agent approvals guide explain which controls apply in each environment.

2. Limit What One Mistake Can Touch

Ask what happens if the agent misunderstands a task or follows an instruction from an untrusted file. Can the same identity delete both the application data and its backups? Can it send a screenshot containing customer information outside the team? A separate VM doesn't help with either problem if you give it the same credentials as your usual workstation.

Keep backups under a separate identity and test that you can restore them. Let the agent run approved checks and prepare changes on a branch; keep publishing, production writes and deployments behind human approval. Enforce those boundaries with permissions. A sentence in a prompt is not an access control.

3. Build a Workflow You Can Repeat

Krainski turns the steps he repeats into skills: reusable instructions for how work should be done in a particular repository. Start with whichever handoff causes the most rework. If pull requests keep arriving without test results, write a verification skill before trying to automate the whole workflow.

StageWhat should come back
Idea, discovery, specA concrete behavior change, constraints and acceptance criteria a reviewer can check.
Plan and reviewA scoped change plan and test strategy. Resolve risky assumptions before implementation.
Implement, verify, recordA branch diff, actual check results and a recording or reproduction steps for the changed behavior.
Pull request and deploymentA handoff that names remaining risks and rollback steps. A human approves merge and deployment.

Krainski reports that one team closed roughly ten times as many PRs per week after four weeks of reorganizing its workflow. That's one team's observation, not a controlled productivity result. On your own team, track review time and regressions alongside the number of merged changes.

The Pull Request Recording

In the GitHub demo at 11:49, Krainski shows how a recording can travel with the pull request. Keep it short and show the action that changed. For a preferences fix, for example, show the initial setting, save a new value, then reload to demonstrate that it persisted. Tell the reviewer which build the clip came from.

GitHub supports video attachments in pull requests. Review the clip before attaching it: GitHub uploads files as soon as you attach them. A recording shows what happened on one path in one build. It cannot tell you whether another user can change those preferences, whether a failed save is handled correctly or whether the controls work with a keyboard. Put those checks in the PR next to the clip.

JQ AI SYSTEMS starter prompt

Prepare evidence for a pull request

Ready to copy

4. Learn From Failed Runs

Start with the evidence available: the instructions, tool calls, diffs and check results. Krainski uses three levels of follow-up, depending on whether something surprised him once or keeps going wrong.

  • Explain one action. If the agent skipped a check, ask what instruction or observed condition it was acting on. Have it explain before it edits anything.
  • Reflect on one run. Compare the plan with what happened, and separate proven causes from hypotheses.
  • Run a retrospective across sessions. Find a recurring failure, change one rule or skill and check whether the next few runs improve.

His formatter example is a good test for your own instructions. A rule about which editing tools to use was interpreted as a reason to skip the standard formatter. Explicitly allowing that formatter fixed the ambiguity. When an agent repeats a mistake, look for two reasonable readings of the rule before you add more rules.

5. Cap Work in Progress

Multiple sessions can do more work in parallel. They also create more decisions, reviews and interruptions for you. Krainski closes with a warning from his own experience: supervising agents can wear you out even when you aren't typing much code.

Try one active session and one queued task. Batch reviews instead of reacting to every update, and add a second active session only when you can keep up with the output. For a first run, pick a small bug or UI change, require a test and a short recording, then ask whether the handoff made review faster or merely gave you more to read.

Video Chapters

TimeTopic
00:00Introduction
01:41Give your agents a dedicated place to work
04:35Mind the blast radius
08:53Build a repeatable workflow from idea to deployment
11:49GitHub demo: Show the feature working
12:51Explain, reflect and run retrospectives
17:00The mental load of working with agents
18:55Recap
Share
X LinkedIn Reddit
Build Yours

Want a system
like this one?

Book a free 30-minute call. We map your situation, identify the highest-impact automation, and figure out if we are a fit.

Book Free 30-min Call