Context Rot: The Major Hurdle in Scaling Agentic Marketing

Intro

Engineering teams have a name for why their AI agents degrade over a long session. They call it “context rot.” While marketing teams are building on the same agents, the term is not discussed much yet.

They will run into it all the same. Context rot is often the first wall a marketing team hits when it moves agents from a pilot, or MVP, into scaled operations.

Capability is commoditizing. Context is the differentiator.

For most marketing use cases, you cannot really go wrong picking any of the major LLM platforms as your foundation. The reasoning gap between them has narrowed to the point where the choice rarely decides the outcome.

The edge has moved to context. What a system knows about your brand, your customers, and your prior decisions is where differentiation now lives, not in which model sits underneath. I made that case in an earlier piece on collective decision intelligence.

But an agent has only a limited window to use that context.

Here is the catch. The window an agent must work in is limited by infrastructure, and the limit is smaller than the spec sheet suggests.

As of June 2026, the leading models have converged at roughly a one million token window, so the advertised size no longer separates them.

Two things shrink the headline. Benchmarks like NVIDIA's RULER show a model reliably uses only 50 - 65% of its window before quality slips. Then the agent's own setup claims much of what is left before any work starts: the system prompt, tool definitions, history, and retrieved data. What remains for the actual task is often around a fifth of the advertised size.

Context rot gets worse at scale.

Context rot sets in when the window fills up. The agent cannot hold everything, so it compresses older context into summaries or drops it altogether. Detail and nuance get lost, and the agent reasons over a thinner version of its own context. Output gets worse even though the model has not changed, and the decline often starts before the window is even full.

It might be overlooked in a pilot. At small scale the agent is fast and accurate. Then you ramp, and as you add audiences, channels, and history, the misses show up at the edges first while the averages still look fine.

A long, complex campaign lifecycle makes it worse, because it runs across many steps. You cannot rely on an agent to carry context cleanly from one step to the next, and every handoff can drop what mattered. More agents do not fix this, since each handoff loses context and each one in a chain lowers overall reliability.

The first wall most teams hit is not capability. It is usable context.

Architecture is the strategy, not a bigger window.

You cannot wait this out. Context windows will not suddenly grow enough to make the problem disappear, and no new model is going to arrive that spends context for free. The work is architectural, and a few simple moves carry most of the weight. They ease the pressure, though the constraint itself does not go away.

• Bound the scope. Give each agent a narrow job, scoped tools, and its own isolated context, so it spends the window on its task and nothing else. Split into multiple agents only when work runs in parallel or a role needs different tools, permissions, or hard boundaries.
• Tier the memory. Keep only the active task in the window, compress recent history into a warm layer, and push the rest to a store the agent retrieves from on demand. This hot, warm, and cold pattern is standard in production by 2026, and the payoff lives in retrieval quality, so track the hit rate before reaching for a bigger window.
• Reference a source of truth. Stop passing context hand to hand. Agents should mirror how marketers already work, off the brief, the brand guidelines, and the customer data platform. They should read from that same system of record.
• Put humans at the gate, not inline. People should not police every step an agent takes. They judge the output against clear criteria, encode the checks that can be made deterministic, and step in on the genuine edge cases.

None of these raise the ceiling. They make sure you spend a scarce budget on signal instead of noise.

What this means for you

The strategy is architecture. The lever inside it is how much usable context each agent spends, which is why the discipline now is controlling consumption rather than maximizing the window. The goal is agentic marketing that can perform and stay reliable as it scales, so the context advantage you built actually pays off instead of rotting.

For a marketing leader, the takeaway is plain. The model is no longer the hard part. The system you design around it is, and that is yours to control today.