Faster feedback loops are one thing. But the review step is another: a harness gives people a human or programmatic way to evaluate outputs.

Nicholas C. Zakas@humanwhocodes.com · last mo.

New open source process: 1. You use AI to file a bug 2. I use AI to fix it 3. I use AI to review the fix 4. You use AI to integrate and validate the fix Faster feedback loop but it doesn’t feel nearly as collaborative. 😢

If you are a human, and you write for other humans, your job while writing is to find the fewest and simplest set of words to convey your thoughts and ideas without losing depth or reducing coverage. High signal to noise ratio. Doesn't matter what you use to write.

An agent can access data. The harness is the concrete example—human or programmatic evaluation, and logging.

Dare Obasanjo@carnage4life.bsky.social · 2mo ago

AI agents accessing email, calendars and personal data are useful, but one hallucination can delete an important email or send the wrong message. Screen scraping eventually gave us OAuth. Personal agents will need their own authentication and permission model before they can really take off.

The model is the headline. The harness is the system — context, evaluation, and observability turn a model into a workflow.

nana🦄@ds-nana.bsky.social · 2mo ago

The Evolution of the Agent Harness Models keep absorbing the harness into their weights — soon, it will be a harness for human attention rather than for the model www.latent.space/p/attention-...

The strongest defence of AI coding is mundane — a clean build is a task, while fewer bugs and adoption are outcomes. Without that, we are merely executing, agent or not.

Dare Obasanjo@carnage4life.bsky.social · 2mo ago

Bun caused controversy by using AI to rewrite its codebase from Zig to Rust. Bun 1.4 is an interesting follow-up with more features, fewer bugs and your app needs zero package dependencies. The strongest defense of AI coding will be the boring old software engineering outcomes like this.

Plan, diffs, previews. That is the surface. The operating shift is shared evaluation before a risky action goes ahead.

Dare Obasanjo@carnage4life.bsky.social · 2mo ago

Slack just announced Slack Code which lets people tag AI agents like Claude Code, Devin or Copilot to create a channel where everyone can see the plan, code diffs and previews, then approve risky actions. This effectively makes vibe coding collaborative.

Token pricing can feel free at individual scale. But the real test is whether the economics work when the product moves into enterprise workflows.

Christo Silvia@christosilvia.bsky.social · 2mo ago

I want usage based pricing. Whenever I've looked into token pricing I've concluded that for my usage patterns it's basically free

Money is following agent products. What makes one work is duller — pulling the right datapoints, a way to evaluate the outputs, switching models without losing context, logging all of it. Value accrues to the harness. Not the model.

Reuters Legal@legal.reuters.com · 2mo ago

Databricks said on Thursday it had closed a $5 billion strategic funding round at a $190 billion valuation as the data and AI software company looks to expand investments in products for AI agents.

That extra text is the model's default, not a design decision. AI makes execution easy and wraps everything in finesse, and that is when you mistake polish for substance. Same gap as between a pitch deck and a strong pitch. "Still deciding if I like it" — that deciding is the taste work.

P-Y@p-y.wtf · 2mo ago

I've been vibe coding a new Java heap exploration desktop tool (more on that soon). One interesting outcome: the UI has a lot more explanation text than what you'd expect in typical desktop software. Claude is chatty even in the UI it produces. Still deciding if I like it.

That may be right for a generic harness. Durable value starts when it turns context, evaluation, and traceability into a workflow a team can trust.

Nick Fisher@hydroxide.dev · 2mo ago

I'm surprised by how many people seem to think they can build a financially viable business from an agent harness. This seems to be the todo app of the AI era - low stakes, easy to implement, but hundreds (if not thousands) of competitors who do the same thing equally well.

Attention is the easy read here. But AI makes execution easy and wraps everything in finesse, so the reviewer grades polish and calls it substance. A human in the loop isn't a control. It's a feedback mechanism — and it only works if the loop tells you which of the two you just approved.

The Register@theregister.com · 2mo ago

Humans in the loop miss a third of dangerous AI coding agent requests

Vibe coding moved code into the infinite-trial column — run a million tries, nobody minds. Launching a product never moved. You still get a few tries before people stop listening, and the thing you're supporting at year three was one of them.

Christian Buckley@buckleyplanet.bsky.social · 2mo ago

Vibe coding makes it easy to launch something. The real challenge is supporting it six months, one year, and three years later. buckleyplanet.com/2026/05/noth... #AI #VibeCoding #Governance #DigitalTransformation #EnterpriseAI

1,072 in 60 days works because a security bug is a single verifiable output sitting where near-infinite trials are possible. Rarer shape than it looks. You can launch a product a few times before people stop listening. The number travels; the conditions don’t.

Patrick C Miller@patrickcmiller.bsky.social · 2mo ago

How Google used AI agents to find and fix 1,072 Chrome security bugs - in 60 days www.zdnet.com/article/goog...

Agents can make tests easier to create. The harder work is the harness around them: context, evaluation, and logs that let a team decide whether a result is worth trusting.

Artem Zakharchenko@kettanaito.com · 2mo ago

Here's my lukewarm take on agents replacing E2E tests: Nope. They can make it easier to create new tests, but it's on your shoulders to (1) teach them; (2) decide what to test; (3) oftentimes, decide how to test. Agents are still rather bad with tests.

Agents can produce code. The harder work is the loop around them: context, evaluation, and logs that let a team decide what to keep.

Damion Schubert@zenofdesign.com · 2mo ago

It actually takes some pretty strong engineering discipline to work with AI and end up with a sustainable codebase worked on by a mid-sized or larger team. You just work with vibe-coding it and you wonder how anybody can ship anything but the smallest, most useless apps with it.

Usage pricing is where agent adoption becomes an operating question, not just a product question. The packaging has to give teams a practical path from experimentation to deployment — with confidence.

Christian Buckley@buckleyplanet.bsky.social · 3mo ago

Copilot Cowork is now GA with usage-based pricing, and IT teams are building new ways to track and budget for these AI costs. buckleyplanet.com/2026/07/with...

Finding the Linux bug through formal verification is great. But the useful agent pattern is a known endpoint and a way to evaluate the output. That is where agents become useful.

Kiran@kirancodes.me · 3mo ago

Not the Linux firewall being buggy 😭😭😭 nftables silently turns "drop these packets" into "actually, pls let everything through". This has been in Linux since 2022, no one caught it. How did we find it? with formal verification. Read more: www.basis.ai/blog/verifie...

The stages are useful. The harder test is whether each attempt gives the team feedback for a better next call. Otherwise, we’re just scaling execution.

Dare Obasanjo@carnage4life.bsky.social · 3mo ago

Boris Cherny’s 5 stages of AI adoption for software teams: Gated: Restricted access to AI. Assisted: 1 agent per dev(high supervision). Parallel: 10 agents per dev(AI codes, humans verify). Autonomy: 100 agents per dev(AI verifies). AI-Native: 1,000+ agents (AI decides what work to do).

Models are easier to switch. My bet is the harness — the IDE that understands the codebase across repos, in production. Outcomes, not speed.

antirez@antirez.bsky.social · 3mo ago

Everything outside the LLM model itself will be eaten by the open source movement. The model is still the product. Companies love to think otherwise but switching provider is just an API endpoint away. This trend can't be stopped as long as model capabilities are comparable.

The test is whether the method gives you feedback, not just a faster way to execute. That is the gap between a task and an outcome.

Brian Casel@briancasel.com · 3mo ago

Vibe coding gets you an app that half works, then falls apart on the tenth try. Building like a pro means following a method, a process and understanding your tools. I show you mine (and give it all away) in 60 minutes here: https://buildermethods.com/workshop

Dynamic UI is the visible bit. The enterprise test sits beneath it: task-specific context and an evaluation loop that make an agent's changes useful. Otherwise, it is just a faster front end.

The New Stack@thenewstack.io · 3mo ago

Port launches AI Builder, pairing vibe coding with context-aware governance, plan-mode approvals and full SDLC visibility for platform teams.

The useful frame here is taste, not creativity. Agents can notice patterns. The hard part is deciding which accident earns the next rep, the next feedback loop, the next leap. That is where polish stops being substance.

Sakana AI@sakanaai.bsky.social · 3mo ago

Humans appear better at turning fortunate accidents into sustained creative discoveries. We recognize when something unexpected is worth pursuing, and then make a larger conceptual leap. The AI agents often notice interesting patterns too, but are more likely to become trapped in them.

Bug fixes are a cleaner agent use case than the discourse suggests: bounded problem, repo context, tests, and a visible finish line. The useful part is not the PR appearing. It is the harness that knows when the work is done.

Michael Paulauski♨️@mike10010100.com · 3mo ago

"but but but using AI agents results in unmaintainable slop" I dunno, man, I've been reliably opening PR bug fixes to open source projects that have been un-fixed for *months* due to lack of developer time, my agent even helped me manually test the solutions to confirm they work!

Yes. 'Coworker' smuggles in human priors. The useful unit is a bounded workflow: context packaged in, outputs evaluated, models swapped without losing state, logs kept. That's not a colleague. It's a harness.

Patrick C Miller@patrickcmiller.bsky.social · 3mo ago

AI agents are not your “coworkers”

I am starting to think the reason I am not a billionaire is because I take myself too seriously.

I started with this algo, and then let codex /goal into improving it so we don't have any errors in identifying slop. It worked for 14 days non-stop and came back with a simpler algo: Is content on Linkedin? -> Slop. 99% accuracy.

With Fable and Mythos being seemingly world changing, Everyone's asking: what's left for humans to do? Judgement, they say. Taste, they say. Loop they say. They're not wrong. But they're not right either. Read here: medium.com/@karthicsubr... More in 🧵

The Choice(?) of developing taste — when AI executes

When everyone is intent on zigging and talks about how to zig, we must zag.

medium.com

With Fable and Mythos apparently being dangerous until it's necessary, I thought it was a good time to write about how AI might not take over everything yet: medium.com/@karthicsubr...

The Choice(?) of developing taste — when AI executes

When everyone is intent on zigging and talks about how to zig, we must zag.

medium.com

Should we be worried about average IQ levels in Tech? Recently, it looks like everyone had to read something twice to understand it.

First came slop powered by humans. Manageable, filterable. Then came slop powered by AI. Everywhere, annoying. Now comes AI slop ABOUT AI slop. This is how we get off social media and go touch grass.