AI Agent Instability Isn't the LLM's Fault — It's the Harness Engineering

This week, a technical long-form piece dissecting the open-source project MyCodeAgent points out: what determines whether an AI Agent works well isn't usually how smart the underlying LLM is — it's how solid that 'engineering skeleton' wrapped around it is. The industry calls this skeleton the Harness (literally 'horse reins'). Below we break down the three key engineering choices the article lays out.

What this is

In the AI Agent (an AI program that autonomously calls tools) context, the Harness refers to the framework's engineering structure: what to do next, how memory is stored, what happens when things go wrong. The three choices are:

Single main loop — abandoning the 'plan first, then execute' multi-layer nesting, and using one loop where the model decides the next step at every turn. The cost: no way to force planning to come first. The benefit: only one state to track, easier to debug.

Immutable state — every change generates a new object; the original stays untouched. When a bug appears, you can 'replay' from any point in time without rerunning the entire flow. A hard requirement for commercial-grade Agents.

Completion-gate feedback — a 'gatekeeper' first judges PASS / UNVERIFIED / FAIL. On FAIL, it tells the model 'where it didn't meet the bar' and lets it keep going. The subtext: take 'judging whether it's done' back from the model and into the code — the model isn't necessarily good at this.

Industry view

The engineering community broadly endorses this approach: easy to debug, easy to replay, able to block the 'model falsely reporting done'. But four risks are being overlooked.

The single loop surrenders priority control over planning, so for complex tasks like multi-file refactors efficiency can drop — the model has to be asked at every step. Immutable state creates a new object each step, and the memory overhead isn't trivial. The completion gate depends on a 'clear pass criterion', which doesn't apply to creative tasks like writing novels or doing design work — forcing it triggers infinite loops. Plus, MyCodeAgent is an open-source project autopsy, not standard practice in mainstream commercial frameworks like LangChain or AutoGen, so survivor bias applies.

What it means for everyone

For enterprise IT procurement: next time a vendor tells you 'we're integrated with the latest GPT/Claude models', don't rush to sign. The right questions are: how is completion determined, how does recovery from errors work, and can state be replayed? That's where commercial reliability lives.

For individual careers: over the next six months to a year, engineers who 'understand both models and Agent engineering' will be scarcer than pure-model specialists. Business users who don't code can use ready-made Agent products, but should keep in mind when the product will 'declare victory and stop on its own'.

For the consumer market: when your AI assistant 'does half the job and says it's done', don't rush to blame the AI — it's usually an engineering problem. We expect a wave of product iterations on exactly this point in the year ahead. Whoever nails 'completion judgment' first eats the word-of-mouth dividend first.