Skip to main content

Case study 01

Small agents. Deterministic gates. Reviewable delivery.

Instead of loading every tool and rule into one agent, Pi routes repeatable work through CLI recipes, gives focused agents only the context they need, and requires isolated validation before QA handoff.

AI-assisted engineering Backend / Platform Verification Developer experience

Fast proof

A Pi-style local control plane keeps agentic delivery small, token-aware, isolated, and reviewable.

CLI-first boundary

Repeatable work stays in commands, recipes, and validators before any agent spends tokens.

Context economy

Agents get only the task, cookbook, repo profile, and optional code-graph context they need — scoped context keeps output cheap to produce and fast to review.

Contained failures

Worktree isolation, sandboxing, validation, and QA keep fast agent output reviewable.

Problem

A single bloated coding agent can move quickly but becomes expensive to review: too much prompt context, too many tools, weak isolation, and no compact proof that validation or QA happened.

What it proves

Jeff can turn agentic development from a single bloated assistant into a controlled local delivery loop with explicit policy, isolation, validation, and QA evidence.

Reliability pattern

Route approved specs into isolated worktrees, then require deterministic validation and QA evidence before delivery is presented as reviewable.

Sanitized evidence

This evidence is a reviewed sanitized snapshot of the delivery model, not raw transcript exposure. It demonstrates AI-assisted delivery reliability through the contract between task intent, isolated execution, deterministic validation, and QA handoff.

The publishable proof is the shape of the system: bounded task inputs enter an isolated work scope, validation records command-level results, and the handoff summarizes what changed, what was checked, and what still needs human review. Private paths, tickets, hostnames, raw logs, screenshots, and assistant transcripts stay out of the public case study.

Privacy notes

This case study uses the Pi PRD as authoring context and publishes only sanitized architecture language. It does not expose private repositories, hostnames, tickets, transcripts, tokens, employers, or clients.

View next case study

Agent Tools Workstation System

AGENT-TOOLS keeps skills, commands, sub-agents, and host policy in one git-backed control plane, so Codex, Claude Code, and OpenCode can share the same operating model across machines.