Context Optimization

Context Optimization

Long-running agents, especially scheduler-managed ones, accumulate context quickly. Agiwo provides two separate semantic optimization mechanisms in addition to token-window compaction:

• context rollback
• tool-result retrospect

Context rollback

Rollback is designed for scheduler wakeups that discover no real progress.

Typical case:

• a root wakes on a periodic timer
• child work is still running
• the root only confirms there is no new information
• the root goes back to waiting

Instead of keeping that empty turn in context, the agent can call sleep_and_wait(..., no_progress=True). The runtime removes the no-progress steps and returns the agent to WAITING.

Tool-result retrospect

Retrospect handles the opposite problem: large tool results that add little long-term value.

The flow is:

1. the runtime detects a large or accumulated tool-result footprint
2. a system notice nudges the agent to call retrospect_tool_result
3. original tool output is offloaded
4. the context keeps a condensed summary instead of the full payload

The original step content is preserved in storage. Condensed forms are recorded in condensed_content and become the preferred representation when history is reloaded.

Why it matters

These mechanisms are part of the self-improvement story:

• less context is wasted on empty loops
• exploratory dead ends can be summarized rather than repeated in full
• long-lived agents stay more focused without losing execution history