Skip to content

Topic

Agent Harness Engineering

The engineering discipline of building the scaffolding—control loop, tools, context, memory—that turns a language model into an agent.

Current stories

buildConfirmed3 publishers

On-prem agents are now a purchasing decision: Perplexity's local Computer needs $4,800 of Nvidia

Portable Computer's harness is real engineering. The bill of materials starts at a $4,800 DGX Spark or a 24GB RTX card, which turns local agents into a hardware order.

Perspective Coverage

3 publishers
Builder
Builder 33%
Operator
Operator 40%
Investor
Investor 27%

Reality

Evidence55
Adoption
Not scored
Hype gap+20
Incentives70
Confidence60
buildOne report1 publisher

Google Cloud's RRSI throttles how many harness edits a self-improving agent may bundle

Automated harness evolution keeps whatever edits raise its own benchmark score, so the harness ends up fitted to the eval set. Google Cloud AI Research answers with five regularizers lifted from supervised learning.

Publishers:dev.to

Reality

Evidence28
Adoption
Not scored
Hype gap+35
Incentives45
Confidence30
buildOne report1 publisher

Deep Agents now swaps in apply_patch the moment you name a Codex model

LangChain's new harness profiles set prompts, tool implementations and tool names per model family. The company measures a 10 to 20 point gain on a tau2-bench subset it curated from tasks frontier models have not saturated.

Publishers:langchain.com

Reality

Evidence45
Adoption15
Hype gap+18
Incentives80
Confidence55
buildOne report1 publisher

Fable 5.1 binds each thinking block to the exact bytes of the prefix that produced it

Any harness that refreshes the date in its system prompt or slips a status line into the transcript between turns now gets a 400 instead of a completion, which makes append-only history the fix rather than a retry loop.

Publishers:dev.to

Reality

Evidence57
Adoption34
Hype gap−12
Incentives51
Confidence55
buildOne report1 publisher

Treating an agent as just a model leaves the tool contracts unwritten

The New Stack's harness argument doubles as an estimating memo, because the schemas, idempotency keys, permission checks and readable traces it lists are the parts that decide whether a demo survives a reworded question.

Publishers:thenewstack.io

Reality

Evidence42
Adoption
Not scored
Hype gap+12
Incentives
Not scored
Confidence45
leadershipOne report1 publisher

Asana's test-suite sprint reprices the work engineering leaders keep deferring

Up to four coding agents finished in roughly a week and a half a test-suite migration Asana had been grinding through since 2022, which makes the queue of someday cleanups less a matter of capacity than of what a leader schedules first.

Publishers:asana.com

Reality

Evidence46
Adoption58
Hype gap+24
Incentives76
Confidence52
buildOne report1 publisher

The AI browser-testing gap is plumbing: one 19-step case, eight runs, eight causes

A vendor postmortem on one 19-step case argues the demo-to-production gap in AI browser testing is harness engineering. Its own worst outage happened inside the harness.

Publishers:dev.to

Reality

Evidence32
Adoption
Not scored
Hype gap+14
Incentives80
Confidence38
buildOne report1 publisher

Same weights, 70 points apart: the ARC-AGI-3 table has stopped being procurement evidence

One model was verified at 30.16 in the official harness and reported at 100.00 in NVIDIA's. Microsoft's Agent Lightning now trains the harness into the weights.

Publishers:dev.to

Reality

Evidence54
Adoption58
Hype gap+28
Incentives66
Confidence48
buildOne report1 publisher

A token budget is not a stop condition: 2,000 lines for a 40-line job

One rate limiter, a ten-million-token allowance and no termination criterion produced a diff 50 times the size of the job. The published fix only inspects the damage after the run ends.

Publishers:dev.to

Reality

Evidence27
Adoption9
Hype gap+42
Incentives84
Confidence58
productOne report1 publisher

Nvidia moved one model from 30% to 100% without changing the model

The harness did the work on ARC-AGI-3, which puts agent reliability on your engineering backlog rather than in your vendor contract.

Publishers:techcrunch.com

Reality

Evidence34
Adoption22
Hype gap+38
Incentives78
Confidence44
buildOne report1 publisher

The agent had permission. It did not have the current task.

A developer edited a task while an agent was still setting up, and found that permission scopes say nothing about stale instructions. The fix is a revision bound at dispatch and checked before every write.

Publishers:dev.to

Reality

Evidence32
Adoption
Not scored
Hype gap−8
Incentives22
Confidence45
buildOne report1 publisher

Your agent's blast radius is decided by the filesystem, not the prompt

Part 6 of dev.to's Harness Engineering series puts the runtime where tool calls execute at the center of agent safety. Its list of things a bounded environment must prevent reads as an infra checklist.

Publishers:dev.to

Reality

Evidence24
Adoption
Not scored
Hype gap+12
Incentives68
Confidence41
buildOne report1 publisher

Agent Memory Is Two Systems, Not One Dial With Scratch On One End

A dev.to series on harness engineering splits persistence into in-task scratch and cross-session recall, and argues that building them as one system produces neither.

Publishers:dev.to

Reality

Evidence28
Adoption
Not scored
Hype gap+16
Incentives62
Confidence47