Summarizer (AI Context)
A summarizer is a component that compresses an agent’s growing conversation history into a shorter summary so the session can continue past the context-window limit.
In plain English
A long-running agent session piles up turns: planning, tool calls, tool outputs, reflections. If you feed all of it back on every turn, you will hit the context-window limit before the task is done. The summarizer is the component that watches the history grow and, at some threshold, compresses older turns into a structured summary that preserves the important bits — decisions, open questions, canonical values — and drops the noise.
Summarization is quietly one of the hardest parts of agent engineering. Compress too eagerly and you lose critical detail. Compress too conservatively and you run out of context. The good summarizers are deliberate — preserve goals, decisions, key references verbatim; compress tool-output noise; never compress the last few turns so the agent retains recent working memory.
Why it matters for Black Box
Black Box's summarizer triggers when the CEO session approaches about 150K tokens of working history. It emits a structured summary with sections for goal, decisions made, artifacts produced, and open questions — then replaces the compressed turns. Sessions can run for many hours without losing coherence because of this loop.
Examples
- Condensing twenty tool-call/result pairs into a one-paragraph "actions taken so far."
- Preserving exact file paths and IDs verbatim in the summary so later turns can reference them.
- Keeping the last five turns uncompressed as a sliding window.