Summarization

This example shows how the smart runtime compacts long or tool-heavy history before the active context window becomes unusable.

Open source: examples/summarization/summarization.ts

Use this when

• the agent can run long enough to accumulate noisy history
• tool outputs are large enough to threaten context budget
• you want to understand summarization before adding retrieval or recovery

What it shows

• summarization enabled in the runtime
• tool-heavy execution that creates context pressure
• summary records written into state

Run it

cd examples
npm run example:summarization

Core code

const agent = createSmartAgent({
	model,
	tools: [echo],
	limits: { maxToolCalls: 5 },
	summarization: { enable: true, maxTokens: 500 },
});

const res = await agent.invoke({
	messages: [{ role: "user", content: "Start a very long session to trigger summarization." }],
});

End-to-end flow

1. The smart runtime is configured with summarization enabled.
2. The token budget is deliberately small to force compaction in a demo-sized run.
3. The agent accumulates enough context to trigger summarization.
4. The runtime rewrites the live context into a more compact form.
5. Execution continues with the summarized context.

Why it matters

This is the example to read when your agent is expected to keep working after many large tool calls instead of failing under context growth.

Look for

• changes in state.summaryRecords
• what gets compacted versus retained live
• how the final answer remains grounded after compaction

Production takeaway

Summarization is not a cosmetic feature. It is what keeps long-running agents operational without making every turn pay the full cost of earlier history.

Expected output

• the run completes with a final assistant answer
• state inspection shows summary-related records once compaction occurs
• the demo triggers summarization earlier than a production configuration would

Common failure modes

• the summarization threshold is too high for the demo input: no compaction occurs and the example looks uneventful
• you expect raw tool recovery from this example alone: that behavior is demonstrated in summarize-context, not here