Agent orchestration

Coordinating multiple AI coding agents through a structured workflow - assigning work, sequencing stages, isolating their changes and reviewing output - rather than running them ad hoc. defract's core: competitors run your agents; defract runs your process.

Agent harness

The scaffolding around a language model that turns it into a working agent - the control loop, tool access, context management, and verification. The common shorthand: agent = model + harness. As frontier models converge, the harness is where capability increasingly comes from.

Harness engineering

The discipline of building that scaffolding: how work is decomposed, where a run pauses for a human, and what gets checked before code ships. When every model is good, harness engineering - not model choice - decides whether agentic output is trustworthy. It is defract's whole thesis: the engine changes monthly; the harness is what's portable.

Lifecycle

defract's gated pipeline that carries a story from idea to shipped: story / scope → design → architecture → implementation → review → release. Each stage is approvable; nothing skips ahead silently.

Parallel agents

Multiple Claude Code agents working at the same time on different stories. The hard part isn't running more of them - it's coordinating and trusting what comes back. See running parallel Claude Code agents.

Git worktree isolation

Each story runs in its own git worktree - a separate working directory on its own branch - so parallel agents and people never overwrite each other on main. Parallel work is isolated by default.

Agent-reviewed merge

Code is reviewed by agents before it merges. Humans approve scope, design and architecture and then see the working feature - you stay above the diff, not buried in it.

Visual design stage

A built-in stage where a feature's design is produced and approved before any code is written - uncommon among agent runners, and a core part of why defract produces less slop.

Memory

A system that learns your codebase and conventions across builds, so agents carry context forward instead of starting cold each task. See keeping context across parallel agents.

Local-first

Your work and PTY transcripts stay on your device. Nothing about your code is sent to defract's servers. See local-first AI coding.

PTY transcript

The pseudo-terminal session log of an agent's run - the full record of what it did. In defract these never leave the device.

Own-Claude economics

You bring your own Claude and pay Anthropic directly for model usage. defract orchestrates the work; it doesn't resell tokens.

Human-in-the-loop

A mode where agents pause at every stage and hand control to you - for high-blast-radius work where you want the wheel at each step.

Scope / story

The unit of work in defract. You approve the scope of a story up front; the lifecycle then carries it to shipped.