The harness beats the model

A new paper gave a name to something practitioners already do without asking permission: agents get better not because their weights change, but because someone rebuilds the structure around them. Four axes — memory, skills, protocols, harness — explain more of this year's reliability gains than any model release did.

Four axes, zero weight updates

A review paper making the rounds this year — "Externalization in LLM Agents," out of Shanghai Jiao Tong University, CMU, and OPPO — makes a claim that should reframe how you spend your next improvement budget. Agents are not getting more reliable because their weights change. They are getting more reliable because someone rebuilds the runtime around them. The paper groups that work into four axes.

• Memory — state that survives past a single context window, not just a bigger one
• Skills — procedural knowledge captured once and invoked on demand, instead of re-derived every session
• Protocols — the structure that governs how an agent reaches tools, data, and other agents
• Harness — the wrapper that turns the other three into something that actually runs unattended

You are probably already building this

Translate the four axes into anything resembling a working agent setup and the paper stops sounding academic. A folder of memory files that carries context across sessions and across agents is the memory axis, done by hand. A library of packaged, reusable procedures — instead of re-explaining the same multi-step task in every prompt — is the skill axis. A protocol layer that gives every agent the same consistent way to reach the same tools, rather than a bespoke integration per agent, is the protocol axis. And a watchdog that checks scheduled jobs on a cron, detects failures, diagnoses the cause, and applies a fix before a human notices anything broke — running quietly on whatever automation platform already hosts the workflows — is the harness axis, in its most literal form.

None of these four things care which model sits behind them. That is the point.

The model layer just got more expensive to bet on

The same week this paper made the rounds, two frontier labs pushed their next flagship releases — GPT-5.6 and Gemini 3.5 Pro — past their expected windows, and prediction markets tracking one launch date watched the odds of an on-time ship collapse from roughly four-in-five to under one-in-five. Capability jumps are not arriving on a predictable cadence anymore, and each one costs more compute and more calendar time than the last.

What the harness axis actually buys you

The clearest evidence that harness engineering is its own discipline, not just operational hygiene, comes from how you verify an agent's work. The obvious approach — run the same task three times and vote on the majority answer — is expensive and only catches errors that show up as disagreement between runs. A newer approach checks the agent's own reasoning chain for internal inconsistency instead, and does it more accurately at a fraction of the token cost. That is not a smarter model talking. It is a smarter check wrapped around the same model.

A selfheal loop is that idea running on a schedule: watch the scheduled jobs, catch the ones that fail, diagnose why before escalating, apply a known fix, and verify the fix actually held — all without a human reading a log first. None of it is glamorous. All of it is the difference between an agent stack that degrades quietly over months and one that keeps working while you are not looking at it.

The audit before your next model swap

Before the next model upgrade lands on your roadmap, run a cheaper audit first. Does anything in your stack remember what happened last week without being told again? Does a failure get diagnosed and fixed before you find out about it secondhand? Do your agents reach tools through one consistent protocol, or a different one-off integration each time? If the honest answer is no to any of those, that is the upgrade to ship first — it is available today, and it does not wait on anyone else's release calendar.