Hivemind Labs · Insights

What we learn building and running AI, written up properly.

Research notes from the systems we build and run, and our thinking on how a business should work with AI: where it belongs, what it costs, and what should stay with people. The numbers carry their intervals, and the results that went against us are in here too.

  1. A bike shop colleague checking paperwork beside a laptop.

    Perspective · · Latest

    Should an AI agent have to earn our trust?

    Earned Autonomy applies the same principle to AI agents as to a new employee: more freedom comes from demonstrated reliability.

    Read more

  2. A curve showing the hidden risk in an AI agent's work shrinking as good actions pile up, crossing a risk appetite of 1 in 20 at 73 good actions.

    Research note ·

    Checks keep AI agents safe. Trust is what makes them fast.

    The majority of businesses that implement AI agents hit the same wall: a person has to review the work, and the reviewing becomes the bottleneck. Often it adds more work than the agent saves. Checks and sign-offs make agents safer, but they never make them faster. The only safe way to speed up is trust, earned on a track record. In this research note we aim to communicate how to approach building trusted AI agents, based on evidence from our own research, including our earlier work on agents in production.

    Read more

  3. A person reviewing a printed document beside a laptop.

    Perspective ·

    When every AI output needs checking, what have you automated?

    AI can produce an answer in seconds. That does not mean the work is finished. The real measure is how much human effort remains before the outcome can be trusted and used.

    Read more

  4. Colleagues reviewing documents and mapping a process around a table.

    Perspective ·

    Build, buy or partner? Start with the work, not the AI

    Before deciding whether to build, buy or partner, the company needs to understand the work it is trying to improve.

    Read more

  5. Detected violations in τ²-bench test conversations on two models, comparing prompt-only, original gates and a corrected post-hoc gate set

    Research note ·

    Prompt less, check more: programmatic rules for reliable customer-service agents

    An AI agent can follow a rule in one reply and break it in the next. We tested code checks on proposed messages and actions, with specific feedback when a check fails. Carefully designed checks reduced detected violations in our tests. Poorly designed checks made agents less effective.

    Read more

  6. Full context versus memory record with tools, per model, with cost per thousand queries

    Research note ·

    Memory over context: accurate, low-cost long-term recall for LLM agents

    A compact record of dated facts, read by a cheap model with three small tools, beats handing the model the whole conversation history. It wins on every model we tested, at six to nineteen times lower cost, and it is the only configuration that never invented an answer.

    Read more

From the work

Every piece starts with a system we built or a decision a client had to make. Nothing here is written from a title.

Evidence over assertion

Research notes are pre-registered and compared paired, with intervals. Differences inside the noise floor are reported as null.

Negative results included

A result that went the wrong way gets the same write-up as one that went right, with the mechanism where we found it.

Plain language

Written for the person who has to decide, not for the person who already agrees. No hype, and no product pitch.

Have a question like these in your own operation?

Everything here comes out of work we have done. The first conversation is about your business, not a product.