recursiv
Research

An open framework for measuring agentic workflows.

Not a black box. Built in the open with the San Diego Supercomputer Center.

What we studyRead the field notes
AI AT WORK, TASK BY TASKILLUSTRATIVEJANMARMAYJULSEPNOWSort documentsMatch paymentsAnswer ticketsReview contractsWrite code BIGGER DOT = MORE OFTEN RIGHTKEEPS UPDATING
Every part of the loop

Each product starts as a research question.

  1. SimulatorPart 1
    The question

    Can we predict how a team of people and agents will perform before it goes live?

    How we test it

    Run the setup on past work, across models, and compare the forecast with what really happens.

    Ask for the paper →
  2. Builder 
    The question

    Which setup works best for this job?

    How we test it

    Change one thing at a time: instructions, routing, model, where people sign off. Score every version on the same work.

    Field note →
  3. WatcherPart 2
    The question

    How does it really perform, and how is it governed, once it runs?

    How we test it

    Score each finished task from outside, against your answer key, with planted errors that show how much the checks catch.

    Ask for the paper →
  4. LooperPart 3
    The question

    Does the team get better round after round, even with the model held fixed?

    How we test it

    Feed the gap between forecast and result back into the setup, then measure again.

    In progress
  5. Context 
    The question

    What should carry over from one week to the next?

    How we test it

    Compare runs with and without what was learned before, and track what each piece of context is worth.

    Field note →
  6. Control 
    The question

    Which changes need a person to approve them?

    How we test it

    Every proposed change must pass fixed gates. The loop can never change its own gates or how it is measured.

    Field note →
What gets measured

Four groups of measures, on every run.

Agents give different answers on different runs, so the same work is run many times and scored each time.

AccuracyFit for purpose, errors that slip through, false alarms.
Coordination costRouting, hand-offs, and the time people spend checking.
ReliabilitySame answer on a rerun, cost per task, speed.
Safety controlsOverrides that work, and where people stay involved.
Why SDSC

Open research, kept neutral.

Publishing with the San Diego Supercomputer Center keeps the work neutral and open. The goal is an open index of AI at work: for each type of task, how often a setup gets it right, what it costs, and where a person still needs to be involved. Because models keep changing, the index keeps updating.

Two lanes, kept separate. SDSC publishes the research. Recursiv applies it with companies, on their own systems and data.

  1. Graded from outsideA different model and team score the work. Nothing grades itself.
  2. Your answer keyScored against your own rules and right answers, which the agent never sees.
  3. Planted errorsKnown mistakes hidden in the work show how much the checks catch. In research.
  4. Fixed gatesThe loop cannot change its own gates or how it is measured.
  5. Open and rerunnableDesigned to be open, so anyone can rerun it on their own data.
Field notes

What we are learning, as we learn it.

Context

An open context layer

A model with no access to your world is expensive autocomplete. Context makes it useful, and it is exactly where lock-in hides.

Read →Simulator

Ants do not hold meetings

The most studied multi-agent systems on earth coordinate without messaging each other at all. Here is why that matters for software, and why coordination structure is something you should measure rather than argue about.

Read →Control

Run it where you can turn it off

Every serious operator eventually asks the same two questions about their agent stack. Who can switch this off, and can I take my data with me?

Read →Builder

The autonomy slider

The question is never whether to keep a human in the loop. It is which loop, on which task, and at what cost.

Read →Watcher

A benchmark is a snapshot. A track record is evidence.

We hire people on what they have done repeatedly. We deploy agents on what they did once, in a demo. That asymmetry is the whole problem.

Read →Context

The context window is not memory

Stuffing more tokens into a prompt is not the same as building a system that knows things. One is a buffer. The other is infrastructure.

Read →
Want the papers? Parts 1 and 2 of the series are available on request.Ask for the papersBook a demo