Haystack 3.2.0: SummarizationCompactor for long Agent context, TokenBudgetHook and stop_run clean exits

Core Background & Industry Pain Points

Long-running Agents inflate context across tool turns: hard truncation drops goals and open work; unbounded growth hits window and cost limits. Production also needs clean budget exits instead of exceptions, plus shorter pipeline assembly. Haystack v3.2.0 (2026-09-24, PyPI haystack-ai 3.2.0) ships compaction, budgets, and fluent builders together.

Architecture Highlights & Internals

Experimental SummarizationCompactor plugs into CompactionHook with four tiers (historical_turns → historical_summaries → current_task_steps → current_task_summaries), summarizing only as much as needed; min_keep_steps keeps the newest steps verbatim; summaries can use a cheap model. Sample knobs: context_window=400_000, compact_at=0.7, compact_to=0.4. Experimental TokenBudgetHook stops before the next LLM call with exit_reason="token_budget_exceeded". Any hook can state.set("stop_run", reason). Pipeline.add_components() / connect_many() enable fluent construction.

Authoritative Benchmarks & Measured Scores

No separate public latency/accuracy leaderboard in the release notes. Quantified knobs in docs/examples: 400_000 context window, 0.7→0.4 compaction ratios, and TokenBudget sample 100_000 tokens. Breaking: Jinja custom_filters serialization requires unsafe=True on load; Toolset + operator removed in favor of lists.

Developer Hands-on Guide

pip install -U haystack-ai==3.2.0. Attach CompactionHook(SummarizationCompactor(...)) and TokenBudgetHook(max_total_tokens=...) on Agent; for custom stops use state.set("stop_run", ...). Load pipelines with custom filters via unsafe=True, or set HAYSTACK_UNSAFE_DESERIALIZATION only on fully trusted deployments. Docs: SummarizationCompactor, TokenBudgetHook.