As Models Learn the Harness, Tool Vendors Must Climb the Stack
The agent stack used to have a clean division of labor: the model generated output, and the harness handled retries, memory, permissions, tool selection, and recovery. In 2026, that split is collapsing. Basic orchestration, generic tool connectivity, and packaged memory are getting absorbed into model SDKs and managed platforms, while some models are also learning parts of the control loop themselves. For tool vendors, the practical risk is blunt: if your product is mostly plumbing, you may wake up as a default feature.
That shift matters because the losers are getting easier to name. Standalone wrappers around standard tools are the most exposed. The survivors will be the vendors who move up into governance, evaluation, cross-provider control, and deep domain workflows that models still do not own.
What used to live outside the model is moving in
A useful way to frame the current stack is the one Arize AI uses: the real question is not whether something calls itself a framework, but which responsibilities it removes from application code. 1 That matters because the center of gravity is moving from custom orchestration toward packaged runtimes that already ship with memory, permissions, tool integration, and recovery.
LangChain’s Managed Deep Agents is a clear example of that packaging trend. Its value proposition is not “a better prompt loop.” It is a bundled agent environment that centralizes infrastructure burden and makes durable execution, context management, and permissions part of the product surface. 2
"Managed Deep Agents targets the brittle handoff between local prototypes and production environments by centralizing the infrastructure burden."
— 1 Minute Signal coverage of LangChain 2
For builders, this is more than a wording change. Once the harness becomes the default product boundary, the old “we just connect the tools” layer loses pricing power. The moat shifts away from generic plumbing and toward judgment, policy, and workflow depth.
The first things to commoditize are the easiest things to copy
The compression is easiest to see in the lower layers. The strongest version of the argument comes from the “Great Compression” thesis: the agent harness layer is being squeezed from above by platform-native capabilities and from below by open protocols. 3
Pulumi’s 2026 survey of agent building says the quiet part out loud: “The 80% of agent tooling everyone was paying their team to reimplement is now table stakes.” Its advice is simple: start with the SDK, and only reach for a framework when you outgrow it. 4
"The 80% of agent tooling everyone was paying their team to reimplement is now table stakes."
— Pulumi Blog 4
That is why the most exposed vendor categories are also the most obvious:
- generic tool wrappers
- basic orchestration glue
- one-size-fits-all memory
- hand-rolled retries and recovery
- commodity connectors that SDKs now bundle by default
Pulumi’s “skills” pattern is a good illustration of the new default. Instead of stuffing a huge tool catalog into context, the agent loads a small description first and only fetches the body when it decides the skill is relevant. 4 That reduces context bloat, but it also reduces the room for third-party frameworks to claim they alone know how to do the basics.
For founders, the implication is uncomfortable but useful: if your product is mostly the thing the model vendor now ships out of the box, you need a new reason to exist.
Internalizing the loop is a different threat
There is a second shift here that is more strategically important than bundling tools: some models are starting to internalize the control loop itself.
AREX 2 is the clearest example in the source set. It is described as a 27B open-weights model designed to internalize retry, reflection, and pacing that previously lived in external agent harnesses. 5
"AREX 2 is a new 27B open-weights model from BAAI designed to internalize the loop logic—retry, reflection, and pacing—that previously lived in external agent harnesses."
— 1 Minute Signal coverage of Sam Witteveen 5
That changes the systems boundary. If the model has learned its own pacing and recovery behavior, the harness is no longer the sole source of truth for retries or long-horizon control. The source explicitly warns that external retry nodes can fight the model’s internal pacing, making some context-wiping habits counterproductive. 5
This is the important ecosystem consequence: the model is no longer just a tool consumer. It is becoming a policy-bearing actor.
The research around tool-use training points the same way. Trajectory2Gym argues that useful supervision is not merely collected from successful trajectories; it is deliberately constructed as learnable interaction experience. 6 In other words, more of the “harness” is moving into the training process itself. The product surface still matters, but the behavior is increasingly shaped upstream.
That does not mean external orchestration disappears. It means the locus of differentiation moves. The old question was “How do I wrap the model so it can use tools?” The newer question is “What should the model already know how to do before it ever hits my runtime?”
What survives when the basic harness disappears
Not every tool vendor is in trouble. The ones that survive are the ones with moats the model is unlikely to absorb quickly.
The clearest defensible categories are:
Governance and policy Identity, permissions, audit trails, approvals, and runtime enforcement are still hard to fake. Kosmoy’s middleware analysis is explicit that once an AI system starts taking actions, the middleware has to reason about tool identity, user identity, delegated permissions, and action auditability. 7
Evaluation and observability If you cannot measure failures, you cannot manage autonomy. The “Great Compression” thesis is explicit that vendor-neutral evaluation and observability remain valuable because they are not just plumbing; they decide who controls the upgrade path and the deprecation schedule. 3
Cross-provider control Portability is no longer automatic. The practical value now lives in systems that let teams switch providers, route by workload, and keep the same policy surface across models. 3 8
Domain-specific workflow depth This is the hardest moat to copy. Gamma’s argument is that product value comes mostly from context engineering and prompt design, not custom model development. 9 The deeper version of that claim is that a workflow-specific product can own the recurring job, not just the model call.
"Do not confuse model power with product viability; your moat is in the workflow, not the API."
— 1 Minute Signal coverage of Greg Isenberg 9
That is the right way to think about the survivors. A generic wrapper can be swallowed. A governance layer can be standardized. But a domain workflow that embeds proprietary judgment, evaluation, and user behavior is much harder to collapse into a native feature.
The market is moving up the stack, not away from it
A common mistake is to read this shift as “agents are becoming simpler, so the ecosystem is shrinking.” The better reading is that the center of value is moving upward.
Pulumi’s survey says some tasks no longer justify full frameworks because SDKs already include the basics. 4 At the same time, the enterprise middleware sources argue that the action layer creates new requirements around policy, observability, and delegation that simple SDKs do not solve. 7 3
That leaves a narrower but more durable opportunity set:
- compliance and governance tooling
- evals and replayable traces
- cross-model routing and fallback
- domain-specific orchestration
- human approval workflows
- stateful, auditable execution environments
The vendors who sell only the first layer of interaction will feel the most pressure. The vendors who help teams decide what should happen, when it should happen, and under whose authority will have more room.
There is also a practical reason for this. The sources keep converging on the same point: models are becoming better at the mechanics, but humans still want control over the consequences. Nylas found trust, control, and failure handling are the main constraints for teams. 10 Temporal’s survey says engineers still need certainty that their delegates have done the job correctly. 11
That is why “more model capability” does not eliminate the need for infrastructure. It just changes which parts of infrastructure can be commoditized.
What founders should do now
If you are building in this market, the question is not whether to ship a harness. You probably should. The question is whether your harness is still the scarce part.
A useful test is this:
- If the model vendor shipped your baseline tomorrow, would your product still matter?
- If retries, memory, and tool use got cheaper natively, what would remain?
- Are you selling implementation, or are you selling control?
- Can you defend your product with governance, evaluation, and workflow depth rather than tool count?
The answer will decide where you belong in the stack.
For many teams, the near-term strategy is hybrid. Use model-native tools where they are cheap and reliable. Keep an external control plane where the workflow is risky, regulated, or expensive to get wrong. That is not a compromise. It is the shape of the market.
The broader takeaway is simple: the harness is no longer a fixed layer outside the model. More of it is moving into SDKs, platforms, and even weights. That leaves tool vendors with one clear job: stop selling the part the model is learning to do, and start owning the parts it still cannot.