The context engineering maturity model

These four failure modes are the dominant patterns that emerge when enterprise teams deploy agents against r eal production data without a mature context layer. Every failure mode described below is a specific, diagnosable architectural gap.

1. Failure mode #1: fragmentation

Enterprise context does not live in one place. It lives across Salesforce, Snowflake, S3, internal operational systems, document stores, event streams, and application state—often with no single system that has a complete or current picture of any given entity.

The answers agents produce are confident and internally coherent, but they’re built on a partial view of reality. This is especially risky in operational contexts, where that limitation often goes unnoticed until a downstream action exposes it.

This fragmentation is the natural consequence of how legacy software systems evolve. Different systems were adopted at different times and for different purposes by different teams. They were designed to interoperate with humans, not with agents.

2. Failure mode #2: opacity

Even when data exists and is reasonably current, agents often cannot navigate it. They can retrieve chunks of text via semantic search, but they cannot traverse relationships between entities, disambiguate between records that are superficially similar but contextually distinct, or determine which records are relevant to the current decision rather than merely keyword-adjacent to the query.

Opacity also manifests as an inability to enforce meaningful context-level access controls. If an agent is given raw database access or unfiltered API access, controlling what it can see is extremely difficult to get right. A semantic layer makes access control a firstclass property of the context architecture rather than an afterthought bolted onto query outputs.

3. Failure mode #3: speed degradation

Latency is a correctness property in agentic systems, not just a performance characteristic. Agents in agentic workflows make many retrieval calls, and any momentary latency can trigger a cascade of issues.

A complex customer service agent might make a dozen tool calls in a single reasoning loop before arriving at a response. At scale, the difference between slow and fast systems affects cost, throughput, and user experience, as well as sheer speed.

Slow retrieval also affects reliability. Retrieval calls that approach timeout thresholds create intermittent failures that are difficult to reproduce and diagnose. Timeout-induced failures look like different problems depending on where in the reasoning loop they occur. This makes slow context infrastructure one of the most insidious failure modes because it degrades reliability in ways that are hard to attribute specifically to the context layer.

4. Failure mode #4: non-accumulation

This failure mode is different in character from the others. Non-accumulation results in a subtler failure: the agent delivers a fixed level of capability regardless of how long the system has been running or how many interactions it has processed.

The consequence is infrastructure that depreciates rather than appreciates. The value of a nonaccumulating context system is roughly constant over time. It’s bounded by the quality of the static retrieval index and the model’s generic reasoning capability.

A compounding context system, in contrast, becomes more valuable the more it is used, because each interaction adds to the memory and personalization signal that makes subsequent interactions more useful.