Why Agents Suddenly Started Working
Something shifted in late 2025. Engineers who had spent months wrestling with brittle, loop-happy agents suddenly found them completing real tasks reliably. The improvement felt discontinuous — like crossing a threshold overnight — but it wasn't a single model release that caused it.
The real answer lives in the relationship between the model and the system wrapped around it. The Latent Space podcast explored this dynamic with researchers and practitioners, and what emerged is a framework we've been working through ourselves as we build production agents: the agent harness matters as much as the model, and understanding how it evolves tells you what to build next.
What an Agent Harness Actually Is
An agent harness is everything that surrounds the model weights — the environment, tool access, memory management, permissions, and guardrails. Without a harness, a language model is a brain in a vat: capable of epistemic reasoning, but unable to act on the world.
The harness gives the model a body. It lets the model perceive information through context, act through tools, persist state through memory, and stay safe through permission boundaries. The model provides the reasoning; the harness provides the reach.
This distinction matters practically. Benchmark data shows that running the same model across different harnesses produces performance spreads of nearly 24 percentage points on identical task sets — with zero changes to the model itself. Half the agent is the harness.
Harness 1.0: The Bolt-On Era
The earliest agent harnesses were scaffolding bolted onto models that weren't ready to use them. The ReAct prompting pattern — reason, act, observe, repeat — described an agentic loop as a prompting technique in late 2022. The loop existed only on paper, external to the model weights.
When AutoGPT and BabyAGI arrived in spring 2023, the harness sprinted ahead of model capability. They handed models full autonomy before models could handle it. The math is unforgiving: at 95% per-step reliability over a 20-step task, you land at roughly 36% average success. Every loop amplifies what the model already has — and brittle models get more brittle under compounding.
The correct response wasn't to give up on agents. It was to pull the harness back. Tools like Cursor and early Copilot put humans back in the orchestration seat, letting models accelerate human decisions rather than replace them. That wasn't a retreat. It was the right engineering call for that moment.
Working on something similar? Talk to our team about your project.
The Curves Cross
Two curves run through agent harness evolution: what the harness asks of the model, and what the model can actually deliver. The gap between them determines agent effectiveness. For most of 2023 and 2024, the harness asked for more than models could give.
The inversion happened at the end of 2024, when the first reasoning models arrived. For the first time, model capability began to outpace harness ambition — creating a capability overhang. Claude Code, launched in February 2025, was the first product built to exploit that overhang deliberately. It abandoned the IDE for the terminal, gave the model direct file and shell access, and replaced per-action human approval with declarative permission rules.
It worked because the timing was right, not because the architecture was new. The team built for the next model's capabilities, not the current one. That timing discipline is something we think about constantly in our own AI agent development work — shipping the right harness for the model you'll be running in six months, not the one you're running today.
Harness 2.0: The Co-Training Era
The current era is defined by a single insight: reinforcement learning has moved inside the harness. Models are now trained inside agentic environments, not just fine-tuned on static datasets. The result is that models begin to absorb harness capabilities directly into their weights.
Auto-compaction is the clearest example. Rather than a harness managing context window overflow externally, models now learn to compact their own context with awareness of their own memory limits. Once that capability is absorbed, the scaffold that implemented it externally can be deleted.
Anthropic's team recently deleted 80% of Claude Code's system prompt — not because the system got simpler, but because the model got smarter. Production by reduction. The same pattern is playing out in reasoning, tool selection, and multi-step planning.
This is the loop that now governs agent harness evolution: train the model in the environment, watch it absorb a harness capability, delete the scaffold, repeat. You can measure your progress by how much you've removed.
For teams exploring this architecture, our AI development services are built around this co-evolution model — designing harnesses that are ready to shed their own weight as models mature.
Harness 3.0: The Attention Interface
Keep deleting everything a model can absorb. Imagine reaching the end of that process. What's left?
The remaining harness capabilities share one property: they are irreducibly human. Permissions, identity, trust, interpretability, and interruption policy cannot be absorbed into model weights without dissolving their purpose. A model that absorbs permissions has eliminated permissions. These are boundaries that must remain external by design.
What survives the absorption cycle is an interface — not between human and model, but between model and human attention. The harness began as a tool humans used to control models. It ends as a tool models use to work with humans.
Tokens are no longer the scarce resource. Synchronous human attention is. The next harness frontier is a policy surface that governs when an agent is allowed to interrupt, which decisions require human approval, and which it can make alone. Think of it as the attention-layer equivalent of AGENTS.md — not instructions for the model on how to navigate your codebase, but instructions for the model on how to navigate your cognitive availability.
We expect every team shipping serious agentic systems to build something like this within the next year. The teams that get it right will run more agents with less friction — not because the models improved, but because the interface to human attention became as engineered as everything else in the stack.
Ready to build? NerdHeadz ships production AI in weeks, not months. Get a free estimate.
The agent harness evolved from a workaround for model limitations into a co-trained system that models are actively absorbing — and what remains after that absorption is the most human part of the stack. Engineers who understand this cycle build harnesses designed to be deleted, and attention interfaces designed to last. That's the discipline defining the next generation of production AI.
“The measure of agent harness maturity is how much of it you get to delete while retaining the same capability.”
