verifiers v1: Composable Tasksets, Harnesses and Runtimes

What it is

Not an agent — the rig for building and measuring them. Agent frameworks normally ship as one bundle; verifiers breaks the environment into three pieces that compose independently:

• Taskset — "defines the work to be done, i.e. the data, tools, and scoring"
• Harness — "the program that solves the task and produces a rollout: a ReAct loop, CLI-based harnesses such as Codex or Terminus 2, or your own agent"
• Runtime — "can be local (as a subprocess or Docker) or in sandboxes, such as Prime Sandboxes"

Named harnesses supported: Codex, Kimi Code, Terminus 2, Mini-SWE-Agent, plus author-built ones. Runtimes: local subprocess or Docker; remote via Prime and Modal sandboxes.

The interception server

Sits between the agent runtime and the inference endpoint, normalising three API dialects — OpenAI Chat Completions (/v1/chat/completions), OpenAI Responses (/v1/responses) and Anthropic Messages (/v1/messages) — while recording "an agent trace on the fly", plus request/response transformation and tool-response interception.

Dialect normalisation is the quiet part that matters: it is what lets the same taskset run against harnesses built for different providers, which is the precondition for comparing them at all.

The trace claim

• v0: "each turn contains prompt-completion pairs leading to a quadratic blowup"
• v1: "each message is unique in a message graph, ensuring the trace size is linear in turns"

O(n²) → O(n) in storage, by storing traces as message DAGs rather than repeated pairs. The stated consequence is that 100+ turn rollouts become practical — which is the regime where long-horizon agent failure actually lives (see the compositional reasoning bottleneck).

Why this rung, and why it matters here

This is the tooling for the experiment this KB keeps saying nobody has run. Harness-Bench had to hand-fix task, budget and evaluation to isolate harness effects. Verifiers makes that separation architectural — any taskset under any compatible harness, with traces recorded identically across dialects.

So the matched-budget test of Prime Agent — same task set, same inference budget, self-improving harness against a plain one — stops being a bespoke research build and becomes a configuration. The same company shipped the claim and the rig that could check it, three weeks apart.

Three corrections to the secondary coverage

Every secondary write-up calls this "Verifiers v1". The primary release note says otherwise, and the differences matter:

1. It is version 0.2.0, released as a preview under the verifiers.v1 namespace — not a 1.0. The universal "v1" framing overstates maturity; full deprecation of the old path is planned only "post-1.0.0".
2. Released 2026-07-10, not 07-13 — that later date is when MarkTechPost covered it.
3. The two numbers most quoted second-hand do not appear in the primary at all — "matched Harbor's performance on the same tasks" and the GLM-4.5-Air ScaleSWE run (two days, six H200 nodes). They may well be true; they are not in the release note, so this KB does not carry them as sourced.

Limitations

• Breaking change: "The legacy code path is now frozen and will not be actively maintained."
• Preview status: the v1 namespace is not the stable API yet.
• No benchmark in the primary. The release is an architecture note — there is no measurement here showing the abstraction costs nothing, and an interception proxy sitting in the request path is exactly the kind of thing that could.
• No licence stated in the release note. Secondary coverage says open source and the repo is public; treat the licence as unverified until read from the repo.
• Vendor-authored, and it is infrastructure Prime Intellect uses for its own training and evaluation — including of its own agent.

Provenance

technical-report, primary (vendor release note). Architectural claims are checkable against the public repo; the performance claims that circulate with it are not in this document.