# Guide, Verify, Solve — Anirban Chatterjee, Sonar

Source: [AI Engineer](https://www.youtube.com/watch?v=03l29gJXpCE)  
Feed7 permalink: https://feed7.dev/p/guide-verify-solve-anirban-chatterjee-sonar-1igfmbm  
Published: 2026-08-09T17:45:13.000Z  
Trust: Source Linked (source_linked)

## Why Included

A Sonar talk argues that coding-agent gains need repeatable verification, citing a three-month productivity spike alongside persistent static-analysis warnings and complexity.

## Source Summary

A GitHub-project study presented by Sonar found an AI-associated productivity increase that faded after about **three months**, while static-analysis warnings and complexity persisted. A separate human study cited in the talk found participants followed correct AI advice **92.7%** of the time and incorrect advice nearly **80%** of the time.

## Practical Implication

Treat verification as part of the agent harness, not a final human glance. The proposed **Guide–Verify–Solve** loop encodes constraints first, runs repeatable computational and LLM-based checks, then gives the agent tools to remediate findings before proceeding.

## Agent-Ready Context

A GitHub-project study presented by Sonar found an AI-associated productivity increase that faded after about **three months**, while static-analysis warnings and complexity persisted. A separate human study cited in the talk found participants followed correct AI advice **92.7%** of the time and incorrect advice nearly **80%** of the time.

Treat verification as part of the agent harness, not a final human glance. The proposed **Guide–Verify–Solve** loop encodes constraints first, runs repeatable computational and LLM-based checks, then gives the agent tools to remediate findings before proceeding.

This was a vendor presentation, and the material does not establish that AI caused the observed project trends. It also provides no comparative results showing that Sonar's specific products outperform other layered test, analysis, and review setups.

## Connected Context

Feed7 judgment across 409 accumulated Signals:

This adds a concrete control loop to warnings that agent throughput can outpace software quality: constraints and verification must shape each iteration, not appear only at final review. The cited persistence of warnings narrows productivity claims to short-term output rather than maintainable progress, while the high compliance with incorrect advice shows why an unaided human glance is an unreliable safety boundary.

- [Harness Engineering is not Enough: Why Software Factories Fail — Dex Horthy, HumanLayer](https://feed7.dev/p/harness-engineering-is-not-enough-why-software-factories-fail-dex-horthy-1tj1amt) — It operationalizes HumanLayer’s maintainability warning by placing constraints and repeatable checks inside the coding loop while retaining human ownership.
- [LLM-as-a-Verifier: A General-Purpose Verification Framework](https://feed7.dev/p/2607-05391v1-10hi3eo) — The LLM verifier could implement part of the proposed verification stage, but the talk’s layered checks avoid treating an LLM judge as sufficient on its own.
- [From Agent Traces to Agent Simulations — Rustem Feyzkhanov, Snorkel AI](https://feed7.dev/p/from-agent-traces-to-agent-simulations-rustem-feyzkhanov-snorkel-ai-0zwlzjq) — Replayable production-derived environments would let teams test whether the Guide–Verify–Solve loop catches representative failures under fixed conditions.
- [Distributed Attacks in Persistent-State AI Control](https://feed7.dev/p/2607-02514v1-02bqscm) — Distributed attacks across multiple pull requests expose a boundary of per-change verification and imply that checks also need persistent, cross-session state.

## Context Map

- Layer: agent
- Domains: coding, security
- Topics: harness-engineering, agent-reliability, agent-evals

## Uncertainty

- This was a vendor presentation, and the material does not establish that AI caused the observed project trends. It also provides no comparative results showing that Sonar's specific products outperform other layered test, analysis, and review setups.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.