COMPOUND V / ANATOMY
A model is not a system. Compound V is the system.
Compound engineering is the practice of building a system around your coding agent so every run makes the next one smarter. Compound V is ProCoders' implementation of that system for Claude Code.
Compound V turns your coding agent into a reviewed, remembering squad that ships real features and compounds over time.
Primitives
The squad and the arena.
Multi-backend squad
Claude plus other backends (Codex and more) on one team. The router picks the right model and backend for each job instead of forcing one model at one price.
Safe parallelism
Jobs with non-overlapping write-paths run at once. Overlap-prone or external-backend jobs get an isolated git worktree, so parallel agents never collide.
Orchestration
Plan, split, dispatch, judge.
Pre-flight scouts
Archaeology, Domain Expert, and Doc Validator gather grounded context from the real codebase, the domain, and current library docs before a single line is written.
Plan and disjoint partition
A written plan is split into non-overlapping write-paths, so the work can run in parallel without two agents touching the same file.
Parallel dispatch
A manifest fans the work across backends and worktrees, batched and fail-fast, with shared foundations built first.
Three-pass review gate
SPEC, QUALITY, and INTEGRATION grader sub-agents gate every merge; an independent cross-model read backs the high-stakes calls.
Memory
What the agent forgets, the repo remembers.
Resumable state machine
A state file survives sessions, so you resume exactly where you stopped and never redo clean work. One feature or a whole multi-feature epic.
Memory and Skills
Skills are procedural memory, the how-we-do-this that carries across projects; project memory persists between runs instead of living in one chat.
Routing lessons
A recorded lesson overrides the default the next time a matching job appears. This is the part that compounds.
Self-improvement
The loop that sharpens.
The compounding loop
Fail, investigate, verify, distill, consult. A failure becomes a checked rule the next run reads instead of re-deriving it from scratch.
A knowledge base that refreshes
V-onboard builds a citation-verified map of the architecture and re-checks it for staleness as the code drifts, so context stays true.
From empty repo to shipped feature
One invocation runs the whole flow, fail-fast and resumable.
The loop that compounds
A failure becomes a rule the next run reads instead of re-deriving.
Step 12 / Safety boundary
Two gates and a human
A deterministic partition and manifest check gates the whole plan before dispatch, and a git-diff scope gate runs after every job and HALTs the run the moment an agent writes outside its allowed paths. Nothing merges on its own: you review and own the merge. Keeping a human at the gate is a deliberate choice, not a missing feature.
Questions
What is compound engineering, or a self-improving agent system?
It is building a system around your coding agent, with memory, review, and recorded lessons, so each run makes the next one smarter. The model stays the same; the system around it is what improves. Compound V is ProCoders' implementation of that system for Claude Code.
Does Compound V run itself on a schedule?
No. Compound V is invoked, not a daemon, and it keeps a human in the loop. Scheduled, laptop-off cloud runs are a Claude Code platform feature you can combine with Compound V, but the plugin itself does not self-trigger.
How is this different from prompting a single agent?
A single prompt fires once and forgets. Compound V decomposes the work, runs a routed squad in parallel with isolation, gates every merge through a three-pass review, and records what it learned so the next run starts smarter.
Do I need Codex or multiple models?
No. Compound V runs Claude-only out of the box. If other backends like Codex are present, the router will use them for the jobs they suit, but a single-model setup is fully supported.
Where does this sit on the AI-native ladder?
Multi-agent orchestration is the L4 (AI-Native System Builder) to L5 (AI Engineering Architect) band of the AI-native developer model. Take the 3-minute quiz to see where you land.
Is it safe to run parallel agents on my repo?
Yes, by construction. Jobs run only on non-overlapping write-paths, overlap-prone or external-backend jobs get an isolated git worktree, a git-diff scope gate HALTs anything that writes out of bounds, and you own every merge.