Harness engineering doesn't mean building your own harness

It means giving the agent enough context to do the right thing

For this week's Agent Factory podcast, I interviewed Ryan Lopopolo, the engineer who coined the term "agentic harness."

What's in the video

• How to Stop Writing Code Manually: Ryan explains how to sculpt smaller feedback loops into long-horizon agentic workflows via harness engineering
• Harness, Deconstructed: Billy Jacobson delivers a code walkthrough of a simple harness, exploring linear and closed-loop styles.
• The Three-Layer Stack: Smitha Kolan breaks down how to use Gemini 3.8 Flash, Antigravity Boost, and the Google Skills GitHub repo to configure memory, tools, and guardrails.

The thing that struck me was how many best practices for scaling software engineering teams also apply to coding agents.

How much code are you writing manually these days? Let me know in the comments!

Comments

[Paul Irolla]

@annthurium the framing I would add to give the agent enough context is that context is not free storage, it is a recurring cost. Everything you hand the agent so it does the right thing is re-sent on every turn of the session, so before adding something, best to assess if you really needs it

[Raknaos]

Answering the closing question directly: almost none of the glue code, and a suspicious amount of the review comments. That matches how I read Ryan's framing — the harness is the thing you tune, not the thing you build.

Where I'd add a caveat is that buying the harness moves the failure mode instead of removing it. Once the scaffold is someone else's, the levers left are context shape and the feedback loop you give it, which are precisely the two things a vendor has no incentive to make legible. The setups that work for me look less like good tool choices and more like a boring loop: small diff, run something, read the output. Do you think that discipline is something a harness can teach, or is it the part that stays human?

[Jo Do]

The feedback loop matters more than who wrote the harness. Good context should include observable completion criteria and a cheap way for the agent to test its own assumptions, not only more repository text. I still write code manually at the boundaries: invariants, destructive operations, and evaluators. Letting the agent handle the middle works much better when those boundaries fail loudly.

[Marko Krsikapa]

Hi. This is Marko.
I'm a software developer and I'm currently working on a project. We have profiles of software engineers of United States.
In fact, while we would like to collaborate with U.S. remote work companies, we lack the language skills required by some of them.
This has caused us difficulties in previous interviews several times.

Your responsibility is to find a remote job.

Securing a remote job involves several steps, such as an introductory call, a technical interview, a final interview, and, in some cases, live coding.

However, we believe you possess the necessary software development knowledge, so we trust you will be able to pass.
We will schedule a meeting with the recruiter and send you the link.

You simply need to attend the meeting and do your best to pass all the steps.
If you pass all stages and receive an offer, you will receive a subsidy of 15–20% of your annual salary.

For example, if your annual salary is $20,000, you will receive $3,000–$4,000 per month, and you will only need to attend daily meetings.

We will give you hourly or fixed before getting job..

We will handle all other matters related to development.

We have extensive experience with this type of collaboration, so we are confident that we can proceed successfully without any issues.
Please let us know your thoughts on this matter, and feel free to reply us if you have any questions.

Thanks.