Mohith Korangi Work with me

All work

Developer toolingSep 2026 – nowPrivate

Local Agentic Harness

The workspace I use to build with AI agents without lowering the bar.

Agents start every session with no memory, and rules written in prose are only advice. This harness puts the rules in files, hooks, scripts and gates, so a one-line idea becomes merged, tested pull requests, and the agents stop only for decisions that are genuinely mine.

Measured

reusable skills, from spec to post-release smoke
11
specialist subagents per project
8
products built on it so far
3

Problem

Fast agents are easy. Agents that don't skip tests, don't leak secrets, don't approve their own specs and don't force-push to main are the hard part, and that only works when the rules can't be skipped.

Architecture

approved green Spec scenario ids Audit claims checked in code Tests first named by scenario Build guarded edits Review code, security, privacy Live QA real Chrome, evidence PR + local CI on the head commit Merge merge commit only Smoke every route after release
How one change moves through the harness
  • Model
  • Service
  • Interface

Each change starts as a spec with scenario ids. An auditor agent checks every claim against the code, and the spec can only be approved by a script that refuses while questions are open. Tests are written first and named by scenario, the builder can't edit them, and one combined review covers code, security and privacy. Live QA in a real browser keeps screenshots as evidence, local CI runs on the exact head commit, and after release a smoke sweep opens every route.

Decisions

  • Rules live in hooks, not prose

    Two blocking hooks per project stop pushes to main, force-pushes, skipped git hooks, reads of .env files, secret-looking text in edits and self-written approval lines. Each rule has its own tests.

  • Coverage is proven, not claimed

    A scenario counts as covered only when a test that actually runs names it. Skipped or focused tests don't count.

  • Generic skills, project overrides

    Eleven skills live once in the workspace and load everywhere, and a project's own version of a skill always wins.

  • Ask the human only what's theirs

    Money, legal promises, accounts, production and brand decisions are batched into one message with a recommendation. Everything else is decided, logged and reported.

Hard parts

  • Honest results

    Skipped steps report as skipped, never as passed, and QA failures are labelled as product, fixture or harness problems.

  • Agents working in parallel

    Each change gets its own worktree, a lock allows one heavy job at a time, and agents never share files.

  • Learning from mistakes

    A retro step turns each real mistake into a written lesson and, where possible, into a hook or a test.

Quality

  • Used to build AskMyChats, PanelCut and ssTitles.
  • Local CI on every pull request runs secret scanning with gitleaks and semgrep.