Use truncation when older exchanges have no bearing on the remaining work. Summarisation suits tasks whose early constraints must survive after the transcript becomes too large. Where verbose tool responses occupy most of the history, reduce those responses before discarding the conversation around them. Choose by whether the agent can still finish the task correctly.
A smaller prompt frees input space. If the reduction also removes the buyer’s exclusion rule, the next comparison can include a supplier the buyer already rejected. Someone then has to catch the mistake and commission another pass. Compare policies against the same unfinished task before assuming that less context means cheaper work.
Compare what each policy removes
Anthropic describes compaction as summarising a conversation and continuing from the summary. Its context-engineering guidance warns that aggressive reduction can lose subtle details needed later.
Microsoft Agent Framework uses compaction more broadly for several policies, including truncation and summarisation. In that terminology, truncation is one form of compaction. When comparing implementations, establish what each one actually retains.
| Policy | Retained material | Appropriate starting point | Failure to inspect |
|---|---|---|---|
| Truncation | Whatever remains after older groups are dropped | Independent requests whose old exchanges are no longer required | An early condition disappears |
| Sliding window | Recent turns or groups, according to implementation | Work explicitly bounded to recent exchanges | A continuing task outlives its window |
| Summarisation | A generated account of older history plus protected recent material | Work with earlier decisions and unresolved dependencies | A precise exception becomes an approximate statement |
| Tool-result reduction | Reduced older tool output, with other history preserved by that policy | Sessions dominated by verbose search or file results | A source detail needed for the next decision is removed |
Tool-result reduction varies by implementation. Collapsing a result into a short marker gives less evidence than preserving a task-specific extract. Inspect what the next model call receives.
Microsoft’s compaction documentation treats tool calls and corresponding results as atomic groups. Preserve the message relationships required by your provider when removing history. An invalid message sequence introduces an API failure before the agent can attempt the remaining work.
Follow an early constraint into the next recommendation
Consider a hypothetical consultancy comparing suppliers. The original request specifies annual pricing and excludes a supplier already rejected by procurement. Research then fills the conversation with product pages and pricing extracts.
The agent now has to recommend a supplier. A sliding window that retains only recent research can lose the exclusion. A summary saying “compare affordable suppliers” preserves the general objective while dropping annual billing. Either route leaves the agent without information required by the original request.
Before reduction, identify the exact constraints the recommendation must honour and the sources needed to verify them. A compact task record should distinguish completed work from open questions, with references that remain retrievable. The agent can then resume at the unresolved point instead of repeating searches it has already completed.
Keep evidence status explicit. “Price checked” in an agent-written summary records what the agent believed; a retrievable pricing source supports checking the price. If that source is unavailable, the agent should identify the gap rather than silently treating the compressed assertion as fresh evidence.
This does not require retaining every research response. An exact exclusion rule and a reference to the relevant pricing page can be more useful to the next decision than pages of unrelated product copy. The retention policy should make that distinction deliberately.
Compare the work completed after reduction
Use a paired replay: save one representative transcript immediately before reduction, then resume copies under each candidate policy. Hold the model and tool access constant, and give each continuation the same next instruction. Retain an unreduced continuation where it fits within the model limit as a reference; check its correctness too.
Inspect whether the final recommendation honours annual pricing and the exclusion. Record repeated searches separately: they expose lost progress even when the final answer is acceptable. Capture the cost of reduction and subsequent usage alongside review time. A shorter context with more retries has shifted cost into another part of the workflow.
Repeat at different points in a long task, including a second reduction of already summarised history. Keep failed trials in the record and run repeated continuations to expose variation. The guide to AI agent reliability metrics explains why a single successful run cannot establish dependable repeated work.
This is a proposed evaluation, not a benchmark we have run. A forbidden supplier recommendation should block release for this task, even if the average answer score looks good. Passing these cases provides bounded evidence about the tested policy, not a universal reliability guarantee.
Check who controls history and where summaries are made
Microsoft documents its compaction framework as experimental and applicable to agents that manage their own in-memory history. Configuring that local strategy does not change service-managed conversation history in Copilot Studio. Establish who owns the history before commissioning custom reduction logic.
The same documentation warns that summarisation sends tool arguments and results to the summariser client. Choosing a cheaper summariser changes which model receives those details. Check that route against the permitted data boundary before using it with company material; the summariser requires the same scrutiny as the main model.
Select a policy using one unfinished task
Start with unnecessary tool output if that is consuming the context. Where earlier state remains important, compare a summary against the current policy using the constraint-retention replay. Agree which lost condition blocks release and carry the test into the PoC-to-production handover.
For independent requests, a bounded window can be sufficient. A long supplier comparison has a different requirement: the original buying conditions must remain available until the recommendation is complete. Those workloads should not inherit the same history policy merely because they use the same model.
Model Operator is a Dubai-based AI product studio founded by Alexander Chrisostomou. Through AI Initiative Consulting, Alexander works with teams on workflow priorities and implementation decisions. Bring one long-running task and the earliest instruction it must still honour to a Model Operator build conversation; use that case to choose the history-reduction policy.