These are fundamentally different concepts that operate at different layers. Here's an accurate comparison:
They are not the same thing.
| Aspect | Claude Code Subagent | Codev Builder |
|---|---|---|
| What it is | A spawned task within the same Claude session | A full Claude Code instance in its own terminal |
| Isolation | Shares context with parent | Isolated git worktree with its own branch |
| Lifetime | Returns result and terminates | Runs until PR is merged |
| Parallelism | Limited by token context | True parallelism - multiple terminals |
| Git access | Same working directory | Own branch, can commit freely |
| Human interaction | None (autonomous task) | Can ask questions, report blocked status |
A Builder is essentially another human-equivalent developer working in parallel, not a helper task.
Also not the same thing.
| Aspect | Claude Code Subagent | Codev Role |
|---|---|---|
| Scope | Single task | Entire session |
| How applied | Task tool spawns it | System prompt loaded at startup |
| Persistence | Ephemeral | Persistent throughout session |
| Purpose | Parallelize specific work | Define persona, responsibilities, constraints |
A Role (Architect, Builder, Consultant) shapes how the agent thinks and operates for an entire session. It's not spawned for a task - it's who the agent is.
Different purposes entirely.
| Aspect | Claude Code Skill | Codev Protocol |
|---|---|---|
| What it is | Slash command that injects context | Multi-phase development methodology |
| Phases | Single action | Multiple stages with defined transitions |
| Human gates | None | Required approvals between phases |
| Artifacts | May produce output | Produces specs, plans, reviews |
| External review | No | Multi-model consultation built in |
A Protocol like SPIR defines a complete development lifecycle:
Specify (human approval) → Plan (human approval) → Implement (with IDE loop) → Review
Skills are more like shortcuts or macros. Protocols are methodologies.
| Codev Concept | What it actually is | NOT equivalent to |
|---|---|---|
| Builder | Full Claude instance in isolated worktree | Subagent |
| Role | Session-wide persona via system prompt | Subagent |
| Protocol | Multi-phase methodology with human gates | Skill |
Codev runs on top of Claude Code. The Architect and Builder roles use Claude Code's tools (Bash, Read, Write, Task, etc.) but add:
- Isolated parallel execution via git worktrees
- Structured workflows with human approval gates
- External consultation with other AI models
- Persistent project tracking across sessions
Think of Claude Code as the engine. Codev is the operating system that orchestrates it for larger software projects.
Codev ships with several protocols for different types of work:
| Protocol | Use For | Key Trait |
|---|---|---|
| SPIR | New features | Full ceremony: Specify → Plan → Implement → Review with human gates |
| ASPIR | Trusted features | Same as SPIR but without human gates on spec/plan — builder runs autonomously |
| AIR | Small features (< 300 LOC) | Lightweight: Implement → Review, no spec/plan artifacts |
| BUGFIX | Bug fixes from GitHub issues | Minimal: fix, test, PR — no spec needed |
| TICK | Amendments to existing specs | Extends an existing SPIR spec with incremental changes |
| MAINTAIN | Code hygiene | Dead code removal, documentation sync, dependency cleanup |
| EXPERIMENT | Research spikes | Hypothesis → Experiment → Conclude |
How to choose: SPIR for anything significant or novel. ASPIR for trusted, well-scoped features. AIR or BUGFIX for small work. TICK to amend an existing spec.
Porch is the protocol orchestrator. It drives SPIR, ASPIR, TICK, and BUGFIX protocols via a state machine — managing phase transitions, human approval gates, and multi-agent consultations automatically.
When you af spawn a builder, porch orchestrates its work:
- Enforces phase order (can't skip from Specify to Implement)
- Runs 3-way consultations at checkpoints
- Blocks at human gates until approved
- Tracks state in
codev/projects/<id>/status.yaml
You can also use porch manually: porch status 42, porch run 42, porch approve 42 spec-approval.
Yes. All protocols (SPIR, TICK, etc.) work in any AI coding assistant — Claude Code, Cursor, Copilot, or any tool that reads CLAUDE.md/AGENTS.md. Just tell your AI: "I want to build X using the SPIR protocol".
Agent Farm adds parallel builder orchestration, a web dashboard, and automated protocol enforcement. It's optional but powerful for larger projects.
Yes. Register your tower with codevos.ai using af tower connect, then access your workspace from any browser. No SSH tunnels or port forwarding needed.
Join the conversation in GitHub Discussions or our Discord community.