cd /news/artificial-intelligence/your-harness-design-is-probably-bad · home topics artificial-intelligence article
[ARTICLE · art-98803] src=datavorous.github.io ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

Your harness design is probably bad

A new formulation for designing AI 'harnesses' classifies task topology by feedback density, search geometry, horizon, evaluation cost, and constraints, then chooses how explicitly search should be controlled on a spectrum from procedural to agent-based. The author, citing AlphaEvolve and Claude Code as examples, argues that procedural harnesses fix the loop while agent-based ones allow adaptation, but both fail at local optima, raising an open question about whether the search representation itself can change mid-run.

read2 min views1 publishedAug 16, 2026

Harness engineering is the new buzzterm, but given its effectiveness, it's hard to deny the need of it. Most of the literature on the internet about harnesses revolve around coding, and sometimes proving math theorems, but often does anyone try to figure out how would one design a harness, given the task is neither of those.

I came up with a formulation to tackle that problem. In short, we try to classify the task topology, then choose how explicitly search should be controlled.

Task topology can be thought of as a vector consisting of feedback density, search geometry, horizon, evaluation cost, constraints. Whereas, search control could be a spectrum, procedure based on one end, agent based on the other, it would say where adaptation is allowed to happen.

Procedural harnesses like AlphaEvolve fix the loop and let a designed evaluator drive selection. Agent based ones like Claude Code let a general purpose agent watch feedback, inspect traces, decide what to try next, more flexible, but it drifts as artifacts pile up. I read about the two "camps" of harnesses in here.

How do we use this taxonomy? Rule of thumb.

Say you're working on to build a harness for scientific discovery where the hypothesis space is huge and eval is sparse, so improving one hypothesis locally can trap the whole search in the wrong representation. What would matter then? Diversity, exploration/exploitation, and letting search history guide what gets tried next, closer to agentic evolution than fixed loops.

Similarly, coding sits at another extreme where dense and cheap feedbacks + hard constraints are available. The right fault is usually findable and verification is near instant, so a procedural loop with local repair does most of the work

A point to note, harnesses like that of Deepmind's Co Scientist sit on sparse feedback and yet it's control is still low. Instead the agents fill roles like generate and reflect without ever touching the process. So procedural doesn't mean rigid and instead should mean that the algorithm is provided by the designer even while the agents inside it move freely.

But both camps would keep failing the same way anyway, if stuck in a local optimum. Fixed rules can't climb out of a bad basin, roaming agents can climb out but might lose the thread doing it. Which is what pushed me past both axes into a third question I don't have an answer to yet, can the search representation itself change mid run?

Something needs to shout: "Stop optimizing this scaffold, try a different chemotype!" That's not a topology property and not a control setting, it's something else I'm still trying to figure out.

I wonder, what if there existed a "harness compiler"?

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @alphaevolve 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/your-harness-design-…] indexed:0 read:2min 2026-08-16 ·