Why can summarisation make an AI agent less reliable?

How summarisation changes the reliability of an AI agent

Summarisation is not a neutral compression step. It changes what later turns can rely on by deciding which details survive, which disappear, and which are reworded into apparently settled fact. In an agent workflow, that can shift the system from “what was actually seen” to “what the summary now says,” which is a much weaker basis for future action.

The core reliability problem is information loss with authority gain. A long context may contain uncertainty, contradictions, or a clearly marked caveat, but the summary often removes the markers that make those limits visible. Once compressed, the result can look cleaner and more certain than the original record, so the agent may continue with a mistaken premise that is harder to challenge later.

That matters because agents do not merely answer questions, they take actions. If a summary collapses the distinction between observation, interpretation, and decision, the agent can reuse a wrong conclusion as if it were verified state. The failure is especially damaging when the summary becomes the only retained record for approvals, side effects, or other steps that should remain auditable.

What goes wrong when context is compacted

Summaries are vulnerable to error because they are selective by design. A faithful summary of an agent conversation must preserve the provenance of key claims, the confidence level of the source, and any unresolved disagreement. When those cues are removed, the model may infer stability where none exists, especially if the summary is later treated as the canonical context for planning or execution.

This is why compaction can create false confidence. The agent may no longer “see” that a value was provisional, that a user instruction was conditional, or that a prior turn contained a mistake. The summary can therefore amplify a single bad detail into a durable operating assumption, which is far more dangerous than leaving the original uncertainty visible.

Summarisation also increases the chance of drift across turns. Each compression pass can subtly reshape wording, and repeated compaction can turn an approximate statement into a hard rule. For workflows that depend on exact parameters, exceptions, or approval history, that kind of semantic drift is a reliability defect, not just a quality issue.

Why summaries should never replace the decision record

A reliable agent keeps a separation between disposable narrative and decision-critical state. Background chatter, intermediate brainstorming, and redundant explanation can be summarised, but approvals, denials, thresholds, commitments, and irreversible actions need durable representation. When those items are compressed into prose, the system can no longer tell whether it is preserving evidence or merely preserving a convenient paraphrase.

That distinction is practical, not academic. If a later turn must justify a payment, change a policy, rotate a credential, or trigger another irreversible action, the agent should depend on a structured record or explicit state, not on a compressed memory of what someone meant. Summaries are useful for continuity, but they are a weak substrate for authority.

For related agent controls, AI Agent Observability, Audit and Incident Response Guide is useful when you need to preserve attribution and action history rather than compress it away. If the workflow depends on exact permissions and approval boundaries, AI Agent Authorisation Guide helps frame where per-action decisions belong instead of being hidden inside summaries. For the broader identity and delegation picture, Agentic AI Identity Guide shows why retained state must distinguish who acted, under what authority, and with what scope.

Risk and Threat Considerations

Summarisation creates a trust boundary problem: the agent begins relying on a transformed record instead of the original evidence. That can expose organisations to silent logic errors, especially when a mistaken summary is treated as authoritative in later planning, approval, or execution steps. The risk grows when summaries are reused across many turns or fed into downstream automation without independent verification.

Failure mechanism: A wrong or incomplete detail is compressed into a clean summary, the provenance signal is lost, and later reasoning treats the summary as settled fact rather than an unverified derivative of prior context.

Impact: The agent can repeat an error, skip a required check, or trigger an irreversible action on the basis of a false but polished memory, making the failure harder to detect and harder to unwind.

Practitioner Guidance

What to verify: Check whether the summariser preserves the difference between source fact, inference, and decision. If those categories are merged, treat the summary as convenience text only, not as a basis for control decisions.

What not to automate: Do not let summaries be the sole record for approvals, exceptions, revocations, or any action that cannot be safely repeated. Keep those items in explicit state or audit logs that can be replayed and reviewed.

Common mistake: Teams often optimise for shorter context windows and forget that compression changes meaning. The reliable pattern is to summarise disposable detail aggressively while preserving irreversible or security-sensitive state verbatim.

Practitioner takeaway: A summary should help an agent remember, but it should never be the thing that authorises action, because compression removes the evidence needed to challenge a bad premise.

Related resources from NHI Mgmt Group

• Why does agentic AI make model identification less reliable?
• Why do stale entitlements make AI-driven detection less reliable?
• Why do AI agents make web analytics less reliable?
• Why do AI-assisted attacks make bot detection less reliable?