The labs will eat memory. They won't eat durability.

For two years the question about AI agents was whether they were smart enough. It was the wrong question. An agent that reasons well for twenty minutes and then loses the thread, crashes, or quietly declares an unfinished job finished is not a worker. It is a demo with a long runtime.

The question that decides whether agents become a workforce is durability: can a piece of work started on Monday still be standing, correct and accountable, on Friday? Underneath it sits the strategic question every startup building agent infrastructure is now asking. When the labs ship the next model, does my product disappear into it? Our answer: memory is being absorbed, durable execution is being commoditised by the clouds, and the companies that survive both are the ones selling neutrality rather than capability.

What actually moved

In November 2025 OpenAI described GPT-5.1-Codex-Max as its "first model natively trained to operate across multiple context windows through a process called compaction." In July 2026 Cognition said its SWE-1.7 model is trained both to write summaries of its own working state and to resume from them. The platforms followed the same direction from the other side. Anthropic launched context editing and a memory tool in September 2025, a compaction API with Claude Opus 4.6 in February 2026, and memory stores for its managed agents that persist across sessions with versioning and an audit trail. Google's Managed Agents compact automatically at about 135,000 tokens. Amazon shipped managed agent memory in Bedrock AgentCore in October 2025.

AGI House's technical primer on long-horizon agents put the consequence bluntly: trained compaction, a summarisation bolt-on in 2024 that became a training objective by 2026, is "the standing threat to every external-memory startup." We agree. In 2024 context management was something you bought. In 2026 it is something the model does, or a parameter you switch on.

Why compaction is not the whole answer

Anthropic's own engineers, writing in November 2025 about agents that run for hours, said plainly that "compaction isn't sufficient." The failures they describe are not forgetting. An agent tries to do too much at once and runs out of road halfway through a feature. A later session sees partial progress and declares the project finished. Their fix was external state: progress files, git history, a written list of what is actually done.

A study published in April 2026 shows how far forgetting is from the main problem. It analysed more than 3,100 agent trajectories across four domains and found that about 72% of failure labels were process-level: misreading the environment, misunderstanding instructions, planning badly, and letting errors compound. The remaining 27.5% were design-level, and even that bucket is mostly not about running out of context. It includes false assumptions and what the authors call catastrophic forgetting, which means a constraint still sitting in context that the agent stops paying attention to. One failed run can carry several labels.

The authors' prescription is better step-by-step planning, checking and repairing plans while the agent runs, and memory that keeps resurfacing long-range constraints. That is model and harness work. It says summarising history is not the bottleneck. It does not say that checkpoints fix agents, and we should not pretend it does: a checkpoint will faithfully save a bad plan.

The other half of durability

Durability has a second meaning that compaction never touches: whether the work survives the infrastructure under it. A checkpoint after each step, a retry when an API times out, a pause that waits days for a human to approve a payment and then resumes where it stopped, a record of every action for the auditor. This is durable execution. It lives in databases, queues and workflow engines, not in weights. It does not make an agent reason better. It makes every failure, whether a crash or a bad step a human has to correct, cheaper to recover from.

Length is what makes that matter. METR's measurements of how long a task an agent can complete have been doubling every 131 days since 2023 and every 89 days since 2024, and its tracker now warns that measurements above 16 hours are unreliable with its current tasks. Toby Ord showed in 2025 that agent success on long tasks is well described by a constant rate of failing during each minute, so success decays exponentially with length. Our own illustration, with the formula at the foot of this piece: an agent that finishes a one-hour step 90% of the time completes a ten-hour task in one pass only about 35% of the time. Restarting from scratch on every failure, a finished ten-hour job costs about 19 hours of agent time on average. With a checkpoint every hour it costs about 11. At twenty hours the gap is about 72 against 22. The model did not get smarter. The work got cheaper to finish.

Everyone sells it now

Our first version called durable execution the half the labs could not absorb, and named managed agents with real checkpoints as the thing that would prove us wrong. That had already happened. In December 2025 AWS launched Lambda durable functions, which "automatically checkpoint progress, suspend execution for up to one year during long-running tasks, and recover from failures," pitched explicitly for AI workflows. Microsoft's durable task extension brings Azure Durable Functions to its Agent Framework, with agents that survive crashes and wait for humans. Vercel ships durable agents whose tool calls and approvals persist through restarts. Cloudflare rebuilt its Workflows engine for agents.

And yet the independent companies are growing. In September 2026 Temporal raised $550 million at a $12.55 billion valuation, reporting a revenue run rate above $250 million, up more than 200% in a year, with OpenAI among its named customers. Much of that usage is not agents, so it is not a pure agent signal. At the early stage the comparison is even: Restate's Series A and Mem0's Series A were both $20 million. Investors are not picking memory or durability. They are backing both.

If every cloud sells durable execution and an independent still triples its revenue, capability is not what customers are paying for. They are paying for something that runs the same way on any cloud and with any model, and that no single vendor can change underneath them.

The invoices agree

Outcome-based pricing shows the same shift from what an agent can do to what it reliably finishes. Intercom charges $0.99 per resolved outcome and, at Fin's average resolution rate of 76%, says a per-conversation price would bill customers for the roughly one in four conversations that never resolve. Zendesk bills $1.50 per automated resolution, often defined by a 72-hour quiet period. Sierra says an unresolved conversation in most cases costs nothing. A resolution that must survive 72 hours is a durability metric with a price on it. Under outcome pricing the vendor eats every failed run, so the vendor now has the strongest incentive to make work survive, whoever supplies the plumbing.

What failure costs is measured less well than the hype suggests. We found no rigorous measurement of supervision hours for production agents. METR found that roughly half of AI-written patches that pass SWE-bench's automated grader would still not be merged by maintainers. Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear value and inadequate risk controls.

Who is exposed

Sort the market by what each company sells, not by which layer it sits in. Capability sold on its own is exposed everywhere. A memory product whose pitch is that your agent forgets less competes with the weights and the platforms. A durable execution product whose pitch is only that your workflow survives a crash competes with Lambda, Azure and Vercel, each a checkbox away from the customer's existing bill.

Neutrality is the defensible position in both layers. Mem0's founder argues that memory is becoming one of the labs' key moats and pitches a memory passport that travels across vendors. Zep calls itself a neutral, multi-cloud alternative to memory bound to one vendor's stack. Temporal runs the same way on every cloud and with every model. Each is selling the absence of lock-in, at the moment the labs and clouds are working hardest to create it.

What would prove us wrong

If customers stop caring about lock-in, by consolidating on one cloud and one model provider, neutrality loses its premium and the independents become features waiting to be bundled. If a lab platform becomes genuinely model-neutral and cloud-neutral, it competes on the independents' own ground. And if models become reliable enough that the per-minute failure rate falls close to zero, the economics of checkpointing shrink with it. None looks close today.

Capability gets absorbed. Neutrality survives. The agent stack is not splitting between memory and durability. It is splitting between what a single vendor can bundle and what a customer needs to run everywhere.

A note on the arithmetic

We treat each hour of work as a step that succeeds with probability p = 0.9, with failures independent from hour to hour. Without checkpoints a failure anywhere means starting over, and the expected agent hours to finish an n-hour task are (p to the power of minus n, minus 1) divided by (1 minus p). With a checkpoint every hour, each hour needs on average 1/p attempts, so the expected cost is n/p. For n = 10 that gives 18.7 and 11.1 hours; for n = 20, 72.3 and 22.2.