Sage-A1.5 / AGENTIC VERIFICATION

On long-horizon problems,
can a model notice it is stuck
and find another way through?

The most common failure of a search agent is not that it cannot search — it is that it keeps retrying a route that leads nowhere.

Watch the demo See Agentic Verification in action
01 / DEMO

One question,
from spinning to solved.

00:00 00:30
02 / WHAT IT REFLECTS ON

The moment it stops,
this is what it writes.

The full reasoning from interaction 20. It takes stock of what is actually confirmed, admits its earlier guesses do not hold, lists what is still missing, and decides to search a different way.

Five things it checks

    All five have to appear together for this to count as real self-reflection — not just the phrase “let me double-check”.

    Interaction 20 · full reasoning

    03 / SIDE BY SIDE

    The same question,
    answered twice.

    Every search and page-read request the model actually issued. The left run changes approach partway through; the right one never does — 47 tool calls, still wrong.

    Changed approach partway

    30 interactions · 29 tool calls

      Final answerMadhumatiMatches the reference answer Madhumalti (spelling variant)
      Same line throughout

      48 interactions · 47 tool calls

        Final answerJacquelineDoes not match the reference answer
        04 / TURNING IT ON

        One line of config,
        on or off.

        The capability needs no extra tool or endpoint. One line in the system prompt turns it on.

        Agentic Verification

        Off, it is an ordinary search loop. On, the model stops by itself once several attempts have produced no progress, and looks for another way through.

        The two steps it adds

        05 / RESULTS

        Performance on
        BrowseComp.

        All 1,266 questions of BrowseComp, same evaluation pipeline.

        Without the capability72.5%BrowseComp · 1,266 questions
        With the capability82.9%BrowseComp · 1,266 questions
        +10.4 points
        06 / WHAT IS NEXT

        After reflecting,
        fold the stretch into one record.

        The longer a task runs, the longer the history gets. Next: after reflecting, the model decides for itself whether the stretch it just walked can safely be folded into a stage record that replaces it.

        In development

        After reflecting, deciding whether to fold the stretch into a record

        Reflection judges whether the conclusions hold and where to go next. Immediately after, the model judges one more thing: whether the stretch from the previous reflection up to this one can be safely folded up. If it can, it writes a stage record — what was done, what came of it, which assumptions were ruled out, what state things are in — and that record plus the reflection replace the original stretch. If it cannot, the stretch stays as it is.

          Coming soon

          AGENTIC VERIFICATION

          When the road runs out,
          stop and take stock, then try another one.

          The self-reflection happens inside the model’s reasoning and costs no extra search. One line in the system prompt turns it on or off.

          Watch again Back to top