The Agent Stack Is Compressing: When Harnesses Stop Being a Product

01 What the Harness Actually Does

Strip the marketing away and an agent harness is a control loop wrapped around a language model. It maintains a plan, breaks the plan into steps, converts steps into tool calls, feeds results back into context, retries when a call fails or a response drifts, and confines every side effect to a sandbox before anything touches production systems. Planning, tool dispatch, context assembly, error recovery, and isolation: those five jobs are the product. The model supplies candidate next actions; the harness decides which candidates run, in what order, with which permissions, and what happens when they fail.

The engineering existed for a blunt reason. Presented with a bare completion endpoint, a model would produce fluent text describing an action rather than the structured call a downstream system could execute. Harness builders wrote schemas, validators, parsers, and state machines that converted intention into execution. They buffered context windows, summarized old turns, and replayed tool outputs so long tasks could survive past a single prompt.

That glue is why agents became usable at all. A raw model plus a shell is a demo; a model plus retries, sandboxes, and durable task state is something a company can bill for. Understanding exactly which responsibilities live in the harness is the prerequisite for judging whether the layer survives the next few product cycles.

02 Why a Middle Layer Appeared

The timing explains the layer. The models that anchored 2023 and early 2024 could describe a plan far more fluently than they could execute one. Tool calls came back malformed, multi-step plans drifted after two or three hops, and context windows overflowed on tasks any human intern would finish without notes. Nothing about that pattern suggested smarter models were imminent, so the practical answer was to build scaffolding around the gaps.

Startups did exactly that. A framework sold deterministic wrappers: strict output schemas, validated tool signatures, retry budgets, checklist-style plan tracking. Customers were not buying intelligence, because the intelligence was the part that kept failing. They were buying the surrounding reliability engineering, the same way early web companies bought load balancers and connection pools rather than faster servers.

That origin matters for what came next. A layer whose value is compensation for missing capability holds only until the capability arrives. From the start, every harness roadmap carried an implicit expiration date tied to the next model generation, and the honest founders in the category said so in their own pitch decks.

03 Absorption From Above

The pressure from above came from the model vendors themselves. Native function calling replaced brittle output parsers. Computer-use interfaces let a model operate a graphical interface directly instead of waiting on a hand-written integration. Context windows grew from thousands of tokens to hundreds of thousands, removing an entire category of context-summarization machinery. Each capability shipped as a model feature was one a framework no longer needed to fake.

The chart below tracks the clearest public proxy for the shift: SWE-bench Verified, where systems attempt real software issues. Reported results climbed from about 4 percent in late 2023 to 49 percent by June 2024, 70 percent by February 2025, and 77 percent by August 2025, as reported at each release.

Interpretation, labeled as such: a model that solves three of four real issues unaided does not need a supervisor rewriting its plan every few steps. Harness features that compensated for failure become dead weight, then liabilities, because they add latency and failure modes the model no longer has. Absorption-from-above is not a rumor; it is a published benchmark curve.

04 Absorption From Below

Pressure from below came from distribution. Cloud platforms bolted agent runtimes onto services enterprises already bought. Editor vendors shipped coding agents as IDE features. Operating-system makers exposed assistants with tool permissions, sandboxing, and scheduling built in. Each of those vendors treats orchestration as a checkbox, priced into something else, and a checkbox is very hard to charge for separately.

Open source consolidated in parallel. Framework abstractions were forked, simplified, and absorbed into lighter runtimes; inference servers grew native tool-loop support; agent scaffolds became starter templates copied in minutes. When the reference implementation is free and good enough, the commercial wrapper around it needs a second story to tell.

Enterprise procurement sealed it. A platform agreement with one vendor, one compliance review, and one bill beat a six-framework stack in nearly every serious evaluation. Pricing pressure flowed downhill too, as the token-cost collapse charted below changed what a margin on orchestration could even mean. Frameworks did not lose on quality; they lost on position.

05 The Economics of Compression

The economics are unforgiving at both ends of the stack. Orchestration is thin: a loop, some prompt assembly, a policy check. When flagship input prices fell from 30 dollars per million tokens in March 2023 to roughly a dime by 2025, a collapse of roughly 300 times, the raw material under any orchestration margin became close to free. Nearly free inputs cannot support a separate tax layered on top of them.

What still carries margin has a moat the model cannot generate. Proprietary context and data: the customer's tickets, telemetry, and domain records no generic model has seen. Evaluation infrastructure: the systems that score whether an agent actually works, needed regardless of which model sits underneath. Vertical workflow depth: clinical documentation, claims processing, logistics, where the difficulty is the process rather than the reasoning. And the audit and compliance surface, where liability, permissions, and records live.

Notice what those survivors share: none of them is the loop itself. The loop is commodity. The durable businesses in the agent stack sell state, evidence, and accountability around the loop.

06 What History Says About Middleware

The middleware era of the 1990s and 2000s rhymes closely. The vendors that survived platform absorption sold the parts the operating system would not eat: transactional state, message durability, identity, and trust. Transaction monitors and messaging systems outlived dozens of prettier layers precisely because they held the data of record, not the glue of the month.

Mapping that lesson onto agents: durable state means task memory, checkpoints, and rollback that outlive any single context window. Evaluation loops mean a proprietary, versioned measure of whether work is actually done correctly. Permissioning means deciding who and what an agent may touch, enforced outside the model. Those are the transaction monitors and message queues of this cycle.

A caution on epistemics: the benchmark and price figures in this piece are measured and sourced; the mapping to middleware is interpretation, argued rather than proven. History does not repeat exactly, and model vendors may yet leave room that platform vendors did not. But the burden of proof now sits with anyone claiming the loop itself is a business.

07 What Builders Should Watch

Three durability tests cut through the noise. First, does your layer hold state the model cannot: task records, checkpoints, and audit trails that persist beyond any context window and any model swap? If deleting your product costs the customer nothing but convenience, you do not have a layer; you have a convenience.

Second, is your evaluation loop proprietary? If the quality claim can be reproduced by anyone running the same public benchmark, the margin is a countdown. Third, the sharpest test: what happens to your fee on the day the next model release ships your feature natively? If the answer is zero, the roadmap was already written, whatever the deck says.

The compression thesis, stated plainly: the harness was a bridge over missing model capability, bridges get built into the road, and the value migrates to state, evaluation, and accountability. Builders should keep the parts that compound and stop defending the loop.