From Code Generation to Code Confidence: Building Quality into the Agentic Workflow

Qodo solutions engineer, Jonathan Klick and field CTO Shlomo Dalezman introduced the Agentic Toolbox: skills, a CLI, and an MCP server that give coding agents the same review engine and codebase context Qodo brings to pull requests. A live install, a review of deliberately broken code, and a candid read on where agentic development goes next.

Topics covered:

AI taking the first pass at review stopped being enough. The original goal was for AI to review code before humans did. As teams leaned harder on agents, they started asking why they were waiting for the PR at all. The local review runs the same engine as a pull request review, not a trimmed-down version. Findings come back grouped by severity, ready for the agent to fix before the branch is ever pushed.

Agents and humans need different reviews. Depth and noise pull in opposite directions. The deeper a review goes, the more false positives it surfaces. The recommendation is to let the local review go as deep as it can, since agents can work through everything, and keep the PR review quiet so humans focus on architecture, risk, and business logic.

Context is organizational, not per-repo. On day one, Qodo ingests org rules, maps the contracts between repositories, and learns from historical PRs. Because searches run on Qodo’s infrastructure, an agent can reach repos that were never cloned locally. Change a schema in one repo, and you know what breaks downstream before implementation starts.

Software factories are a spectrum, not a switch. Shlomo compared it to self-driving (L1 through L5) and recommended a sequence: local reviews first, codebase wisdom for planning and specs, then the same tools in the factory itself.