---
type: "post"
id: "veriharness-scaling-agentic-verification-for-long-horizon-tasks-ec554a"
slug: "veriharness-scaling-agentic-verification-for-long-horizon-tasks-ec554a"
title: "VeriHarness — Scaling Agentic Verification for Long-Horizon Tasks"
description: "Thesis: Repeated agent attempts become more useful when verification tests both disagreement and shared assumptions instead of treating consensus as proof."
retrieval_nugget: "Thesis: Repeated agent attempts become more useful when verification tests both disagreement and shared assumptions instead of treating consensus as proof."
published_at: "2026-10-07"
updated_at: "2026-10-07"
record_date: "2026-10-07"
date_kind: "published_at"
topics: []
source_urls: ["https://veriharness.com/"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:veriharness-scaling-agentic-verification-for-long-horizon-tasks-ec554a"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:e4cba34710d4889ecf167e9958ef250f8c327463b543e0f4eb6f5f75de259750","reviewed_at":"2026-10-07T20:07:12.855Z","source_evidence_count":1,"verified_claim_count":2,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":2,"implication_count":2,"watch_condition_count":2,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"Repeated agent attempts become more useful when verification tests both disagreement and shared assumptions instead of treating consensus as proof.","observed_facts":[{"text":"The project page says VeriHarness uses the same model for candidate rollouts and verification, with separate investigation contexts for disputed and unanimous claims.","source_urls":["https://veriharness.com/"]},{"text":"The project reports a Gemini 3.5 Flash average native benchmark score increase from 47.2 to 53.4 across five benchmarks after evidence-backed revision.","source_urls":["https://veriharness.com/"]}],"mechanism":"Two independent investigations use workspace evidence to resolve competing claims and challenge agreed claims; a fresh adjudication context combines their evidence records to select and revise an artifact.","why_now":"The project page presents a released harness and evaluation results for long-horizon work across documents, spreadsheets, code, and multi-file tasks, making verification architecture a concrete agent-design question rather than a purely conceptual one.","implications":["Agent quality can depend on evidence gathering and structured verification, not only on the base model or number of attempts.","A verification record can make revisions more inspectable, while the extra investigations impose tool-call and latency costs."],"evidence_boundary":"The reported results concern five benchmarks, ten rollouts per task, and two frontier models; they do not establish performance across all agents, tasks, or deployment conditions. The page notes that verification adds tool calls and latency.","watch_conditions":["Whether independent evaluations reproduce the reported gains on new models and task suites.","Whether the added verification cost is acceptable for professional workflows."],"related_records":[],"new_branch_reason":"No prior New Runtime record is supplied; this begins a distinct record about the particular verification design and its reported evaluation."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/veriharness-scaling-agentic-verification-for-long-horizon-tasks-ec554a/","markdown":"https://newruntime.com/posts/veriharness-scaling-agentic-verification-for-long-horizon-tasks-ec554a.md","json":"https://newruntime.com/posts/veriharness-scaling-agentic-verification-for-long-horizon-tasks-ec554a.json"}
---

# VeriHarness — Scaling Agentic Verification for Long-Horizon Tasks

## Retrieval answer

Thesis: Repeated agent attempts become more useful when verification tests both disagreement and shared assumptions instead of treating consensus as proof.

Thesis: Repeated agent attempts become more useful when verification tests both disagreement and shared assumptions instead of treating consensus as proof.

Observed facts: VeriHarness uses the same model for candidate rollouts and verification, with separate investigation contexts for disputed claims and unanimous claims. The project reports a Gemini 3.5 Flash average native benchmark score increase from 47.2 to 53.4 across five benchmarks after evidence-backed revision.

Mechanism: Two independent investigations use workspace evidence to resolve competing claims and challenge agreed claims; a fresh adjudication context combines their evidence records to select and revise an artifact. This creates a checkable path from candidate outputs to a delivered result, rather than relying on majority vote alone.

Why now: The project page presents a released harness and evaluation results for long-horizon work across documents, spreadsheets, code, and multi-file tasks, making verification architecture a concrete agent-design question rather than a purely conceptual one.

Implications:
- Agent quality can depend on evidence gathering and structured verification, not only on the base model or number of attempts.
- A verification record can make revisions more inspectable, while the extra investigations impose tool-call and latency costs.

Evidence boundary: The reported results concern five benchmarks, ten rollouts per task, and two frontier models; they do not establish performance across all agents, tasks, or deployment conditions. The page notes that verification adds tool calls and latency.

Watch conditions:
- Whether independent evaluations reproduce the reported gains on new models and task suites.
- Whether the added verification cost is acceptable for professional workflows.

This starts a distinct analytical record because no prior New Runtime record is supplied; the available material supports evaluating this particular verification design without assuming a related prior item.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)