0:01You give an AI coding agent a feature to implement and it starts working through the codebase.
0:06It reads the files, plans the steps, and starts implementing.
0:10It even writes its own tests.
0:12All tests passed. Task complete.
0:14But look at what those tests actually check.
0:17A result exists. A mock was called. A value has the right type.
0:21Everything looks verified. But would you ship this in production?
0:25Why does this still feel unsafe?
0:29Agents are good enough now to do real work. But that does not mean they will get it right on their own. So the question is: how do you make their work reliable?
0:38That's harness engineering. Control, context, and checks. And that's what this video is about.
0:46Let's break those down.
0:47Control. What the agent is allowed to touch, and where it runs.
0:51Context. What it sees, and how much of that is actually useful.
0:56And checks. What has to be true before the work counts.
1:01Let's start with "Control".
1:02In the last video, we said a raw model is like a junior developer with root access and no code review. Control is how you take that away.
1:10Every task gets its own environment. Fresh machine, scoped credentials.
1:15The agent does its work inside the box. Nothing that touches production.
1:19External checks run. The agent can't skip them.
1:22And when it's done, the whole environment is destroyed. Nothing persists. Nothing carries forward.
1:27Stripe's Minions team runs this at production scale. Thirteen hundred pull requests a week, every one in its own disposable environment. You can only let agents work at that scale when a wrong move can't hurt anything.
1:39Here's the key thing to differentiate Control from just regular Prompting. Prompting says "please don't." Control says "you can't."
1:49Now let's talk about context, what the agent sees.
1:52When it comes to context, more is not better. The right context is better.
1:57Without guidance, the agent explores on its own. It pulls in files that look relevant but aren't. And then it makes decisions based on the wrong examples.
2:05Only a handful of files actually matter for this task. Everything else isn't just wasted space, it's actively misleading.
2:12This is context pollution. The wrong files don't just take up tokens, they cause wrong decisions.
2:20So how do you give the agent the right context?
2:22One pattern is "progressive disclosure". Instead of loading everything upfront, you let the agent pull in context as it needs it.
2:30Let's see an example.
2:31Here's a "skills" catalog available to the agent. Each entry has a name and a short description. The full content isn't loaded yet, just the descriptions.
2:40If the agent needs to verify its work. It scans the descriptions, and selects the one that matches.
2:46Only now does the full skill file load. Steps, references, configuration.
2:52All the depth was there, but the agent didn't need it until this moment.
2:57The other pattern is ambient context, context the agent should always have, before any specific task begins.
3:04Without that orientation, the agent guesses. It runs the wrong test command, because nothing told it the project uses a different one.
3:11Same agent, same repo. But now there's a small file loaded at the start. Right command. Tests pass.
3:18This is the entire file. Five or six lines, and the agent reads all of it every turn.
3:23It doesn't add complexity, it removes bad decisions.
3:28So you've got two patterns now, and they're opposites.
3:31Progressive disclosure, capabilities the agent might need, loaded when the task calls for them.
3:36But ambient context, rules that shape every decision, those are present before the task begins.
3:42Same agent, same project, two different loading strategies because the information is fundamentally different.
3:49And finally, checks. What has to be true before the work counts.
3:54Why not let the agents write its own tests? Same reason why you dont let a student grade their own exams.
3:59Agent writes a bunch of tests. All pass. You get verification theatre.
4:03Now make the agent run tests were written by someone who already knew what correct looks like. Edge cases, real scenarios, what the real-world messiness looks like.
4:12The agent fails against a real standard, has to go back and fix it, and only then does it pass. Independent verification changes the outcome completely.
4:23So, what does good agentic verification look like?
4:26You make it go through an independent pre-completion gate.
4:29The spec is a domain expert who knew what the output had to be.
4:32The linter is the team's conventions.
4:34The required checks are the CI pipeline.
4:37And the retry threshold, that's the system designer saying if you've tried this five times, you're stuck. Just stop and escalate.
4:44The agent isn't being graded by its own rules.
4:47It's being graded by the people who know what good looks like, even though they're not in the room.
4:52The writer doesn't grade the exam. The system does. That's checks.
4:57So we have covered control, context, checks.
5:01Once those are in place, a harder set of questions opens up.
5:04What exactly should the agent build? How do you decompose the work?
5:08How should it reason through the task? What's the right level of autonomy?
5:12And how do "you" decide what "correct" actually means, when the output is complex?
5:17That's where the next part of this series goes. See you there!

Harness Engineering: Control, Context, and Checks

Created by Abhyuday DesaiUpdated 2 Oct, v3Runs 5 min 26 sec

0 views

About this video

Part 2 of the Harness Engineering series. Covers the three pillars that make AI agent output reliable: control (sandboxed environments, scoped credentials), context (progressive disclosure and ambient context patterns), and checks (independent verification gates). Includes real examples from Stripe's Minions team.

What changed

v302 Oct 2026abhydesai5 min 26 sec

Re-rendered in the studio with chapters and a transcript. The narration is now by Alex, the series' voice; v1 was narrated by Kyle.

v1 has no transcript to compare. 29 sec shorter.

0 views0 likes on this version

v202 Oct 2026Withdrawn, no longer available to watch
v113 Apr 2026abhydesai5 min 55 sec

First published cut.

0 views0 likes on this version

Watch v1