Soft degradation can bind before the hard cap even when required evidence fits

Type: kb/types/note.md · Tags: learning-theory, foundations, deploy-time-learning

A useful model of agent context has two boundaries. The hard token limit is the maximum input the model accepts; crossing it returns an API error. The soft degradation boundary is the point at which task performance drops through missed instructions, shallow reasoning, or unused context even though the output remains fluent and well-formed.

For quality-sensitive agent work whose required evidence still fits inside the provider window, the soft boundary can bind first. The evidence here establishes that possibility across several task shapes, not how frequently it is the first constraint in representative workloads. Some tasks really are hard-cap-bound, and no prevalence estimate in the cited studies licenses “often.”

Dimensions of the soft bound

At least three distinguishable pressures can reduce usable context before the hard cap: volume, relevance/interference, and complexity. They overlap, and prompt framing or information arrangement can shift several at once. This taxonomy groups observed failure pressures; it does not assume they all share one internal cause.

Volume

More tokens can make it harder for the model to recover and use the right material. The “lost in the middle” result (Liu et al., 2023) shows primacy and recency bias in long-context retrieval: models underuse information in the middle of a sequence. Agent prompts inherit that risk whenever they require recovery from long, weakly scoped input. Across eleven tested models and four synthetic retrieval, aggregation, and sorting question types, Paulsen reports that measured Maximum Effective Context Windows fell well below providers' maximum accepted windows and shifted with question type, with some measured gaps exceeding 99%. The study's simple generated records do not represent ordinary agent workloads.

Relevance/interference

Not all tokens cost the same. Irrelevant context can do more than add volume; it can actively interfere with task execution. In GSM-DC's controlled synthetic math problems, increasing injected irrelevant context reduced reasoning accuracy across six tested instruction models; error grew roughly as a power law in distractor count with a steeper exponent at greater reasoning depth, and the disruption affected both correct path selection and arithmetic execution. The measured rate is bounded to the benchmark's templated problems, distractor range, depths, and models.

The same pattern appears in an agent benchmark. Inserting irrelevant task sequences between dependent subtasks to create 25,000–150,000-token web-agent histories reduced four tested models' success from roughly 40–50% in baseline conditions to below 10% in long-context conditions. Loops and loss of the original objective were prominent, while task-relevant summary retrieval produced only modest improvement. Because the benchmark changes history length and intervening task content together, it does not isolate volume from interference or show that agents treated stale history as live state.

Complexity

Some context becomes expensive not because it is long, but because it is hard to interpret or compose. A reference that the model must resolve adds interpretation work, and deeper compositional structure may impose a similar burden. On ConvexBench's deeply composed symbolic-function tasks, one-shot reasoning fell from F1 1.0 at depth 2 to about 0.2 at depth 100 even though the depth-100 input was 5,331 tokens; agentic reasoning with focused context reached F1 1.0 across the evaluated depths. Token count alone therefore does not predict usable capacity on this benchmark. What remains open is whether its one-shot failure is specifically a context-management limit, a missing reasoning procedure, or some mixture of both.

Open questions

The main unresolved question is interaction. GSM-DC shows that distractor count and reasoning depth interact in synthetic math problems. The web-agent benchmark instead shows agent-level degradation in long, dependent multi-session histories where length and intervening task content vary together; it does not isolate their interaction. We still do not know how stable the GSM-DC interaction is across natural-language tasks, partially relevant material, or model families.

Candidate mechanism: workspace saturation and displacement

One candidate explanation is that flexible task computation depends on a limited internal workspace. The J-space experiments identify a selectively engaged, verbalizable subspace that appears to broadcast intermediate representations. Ablation disproportionately harms flexible, multi-step tasks while leaving much automatic computation intact. That pattern is consistent with volume, interference, and dependency depth competing for representations needed by the active task.

This evidence does not yet establish J-space as a general mechanism for long-context degradation. The study does not vary long-context workload against workspace occupancy. Its method captures only part of internal representation. Some computation bypasses the measured space, and the process that admits material into the workspace remains unexplained. Still, the hypothesis makes a useful prediction: matched increases in volume, interference, or compositional depth should degrade performance when they increase competition, displacement, or dependency load in the operative workspace. A degradation regime with no corresponding workspace effect would count against it.

The soft bound is invisible

Crossing the hard limit returns an error. Crossing the soft boundary does not. The model can remain fluent while missing instructions, following a stale objective, or leaving relevant context unused. The failure becomes visible only in downstream task quality.

The soft boundary is also not one stable number. It shifts with task type, dependency depth, relevance mix, arrangement, prompt framing, and model version. A provider's window size therefore describes acceptance capacity, not the usable capacity of a specific workload.

Consequences

Treat advertised window size as an upper bound, not a usable budget. The operative budget depends on the task, arrangement, and model version.

Reduce interference and complexity before they enter the call. Selective loading and scoped state determine what belongs in the active problem frame; summarization only compresses what was already selected. Decomposition and externalized state reduce dependency load. Frontloading removes work from the consuming call by pre-computing instruction parts whose inputs are already known. These are rational responses to a boundary that cannot be measured once for all tasks, as other soft-bound traditions also suggest.

Exploit high input control despite low processing observability. An orchestrator can choose every token that enters a call even though it cannot directly observe how effectively the model used them. Default-loading session history squanders that control; session history should not be the default next context. The heaviest-fork feasibility note extends the same concern to work split across sub-agents.

Relevant Notes: