Skip to content
Primary sourcePodcastHarness longreadPrevious reviewSeries startCompression costAnthropic: context
Research Insights Made Simple
Research reviewResearch Insights Made Simple

What should the agent see next?

Goals, evidence selection and reacquisition through a CI example

/ Research Insights Made Simple · Agent harnesses 19/19 · Next context

Slide contents

  1. 1. What should the agent see next?

    Goals, evidence selection and reacquisition through a CI example

  2. 2. Context serves the next step

    Goal, evidence retention, reacquisition and CI verification

  3. 3. Context already belonged to the harness

    Deepening evidence selection after tools, memory and coordination

  4. 4. One project needs different views

    Atlas: diagnose code or confirm an external run

  5. 5. GACC tests retention of relevant evidence

    Preprint dated October 15, 2026

  6. 6. The goal precedes specific questions

    One SymPy trace: repair validation or algorithm behavior

  7. 7. One view serves several hidden questions

    Compressor, reader and judge receive different inputs

  8. 8. One omission breaks complete goal coverage

    Micro counts answers; Macro requires every answer for a goal

  9. 9. Answering everything is much harder

    Full benchmark · retention budget: 10% of source context

  10. 10. A mismatched goal loses useful evidence

    GPT-6.1-Sol · testmini · 10% retention budget

  11. 11. Budget limits size, not quality

    The retained view is capped at 10% of source length

  12. 12. This tests compression, not whole agents

    Static answerability does not replace task execution

  13. 13. Dropped state may need reacquisition

    Shuyu Liu · arXiv:2608.16370v1 · August 17, 2026

  14. 14. Completion can hide extra retrieval

    GPT-5.5 · High-IR · full history versus sliding 5×

  15. 15. Restoring state reduces repeated retrieval

    DeepSeek · High-IR · separate intervention batch

  16. 16. Cost depends on model and environment

    Extra retrieval did not appear in every probe

  17. 17. References let agents fetch details later

    Engineering guidance: initial anchors plus retrieval when needed

  18. 18. A context view starts with evidence

    For c3 diagnosis, select evidence that can be checked

  19. 19. Compaction retains decisions and uncertainty

    Shorten history while preserving access to its evidence

  20. 20. Findings travel with their evidence

    The coordinator receives a claim, evidence and limits

  21. 21. Log text does not grant authority

    Preserve provenance and distinguish data from instructions

  22. 22. Evaluate outcomes and reacquisition cost

    Back to episode 42: outcome, trajectory and comparison conditions

  23. 23. Good context can be tested through action

    Sufficient evidence for the next goal, with explicit gaps