Plan with your coding agent, split independent work, and coordinate native workers inside your existing Claude Code, Codex, OpenCode or Pi session.
Preferred workspace: Orca. Use the Orca setup and execution plan with one coordinator and one scheduling layer. Start with native subagents inside Claude Code/Codex; use Orca-supervised worktrees when you need separate visible workers or mixed agents.
The plugin owns the workflow: clarify → agree interfaces → plan dependencies → run independent slices → integrate → verify. Your host owns the model, credentials, tools and session. Planning decisions stay with you. There is no external agent launcher in the plugin path.
Use multiagentic-engineering to plan this feature with me. Inspect the code first, identify the decisions I need to make, then propose independent slices with file ownership, dependencies and acceptance checks. Keep worker context small. Wait for me to approve the plan before implementation.
After reviewing the plan:
Execute this plan using native agents. Keep shared files under one owner and verify the combined change before calling it done.
Ordinary substantial coding requests can also trigger the installed skill automatically; selection depends on the host/model. Explicit invocation is the reliable entry point. Small or tightly coupled work stays with one agent when parallelism adds overhead.
These are one-time package installation steps, not a runtime agent-spawning CLI. See host-specific setup and migration for app support, preview-branch installation and capability limits.
| Host | Package / entry point | Worker execution |
|---|---|---|
| Orca | Underlying agent plugin/skills; workspace profile | Native subagents or Orca-supervised tasks, one scheduler |
| Claude Code | Marketplace plugin; /multiagentic-engineering:multiagentic-engineering |
Native Agent tool, worker/reviewer definitions |
| Codex | Repo marketplace plugin; $multiagentic-engineering |
Native collaboration tools when exposed |
| OpenCode | Native JavaScript plugin; /mae |
Native Task/subagent tools |
| Pi 1.4 SDK | Pi package; /mae |
In-process SDK child sessions via mae_dispatch |
| Other Agent Skills hosts | Copy/install the shared skills | Native delegation if available; otherwise sequential |
Claude Code, in its session:
/plugin marketplace add Chagatai404/multiagentic-engineering
/plugin install multiagentic-engineering@multiagentic-engineering
Codex: add this repository as a marketplace using its plugin browser or:
codex plugin marketplace add Chagatai404/multiagentic-engineeringThen install multiagentic-engineering from that marketplace in /plugins or the
app's Plugins Directory. This is a repository plugin, not a published universal listing.
If your surface does not support local/repo marketplaces, use the skill-only option.
Pi:
pi install git:github.com/Chagatai404/multiagentic-engineeringFor OpenCode, clone the repo, run npm install --omit=dev --ignore-scripts inside it,
and add the absolute file:///.../adapters/opencode/index.mjs URL to the plugin
array in your existing opencode.json. Preserve other settings. The plugin registers
its skills, /mae, mae_plan and worker/reviewer agents in memory.
The agent writes a decision-complete .plans/<topic>.md and a compact JSON dispatch
plan. Unresolved decisions block dispatch. Each slice delivers a narrow verifiable
behavior across the necessary layers and specifies objective, read pointers, exclusive write ownership,
interfaces, dependencies, acceptance and checks. The validator rejects cycles,
unknown dependencies and unordered write/write or read/write overlap.
Only verified dependencies unlock a consumer. A worker report is not verification. The coordinator integrates serially, checks real changes, and runs acceptance checks. Plan changes invalidate saved dispatch state. Interrupted work is inspected before retry.
Default example: two workers, at most three, a 6,000-character packet, a 1,200-character handoff and one attempt per worker. These are adjustable context/attempt limits, not hard token or spending caps. Worker context excludes sibling transcripts and unrelated skills. Parallelism can reduce elapsed time while increasing total tokens; no measured savings are claimed yet.
The shared interview combines JMO's project
spec interview with Matt Pocock's decision-tree
questioning, terminology and design decisions. Matt's vertical slicing, behavior tests,
and spec/standards review also shape the engineering workflow. ste retains JMO's
clear technical writing guide. See source selection and adaptations.
Optional UI work can use its better-ui and animate skills separately. We do not load
them into every engineering worker. Its Codex CLI handoff is excluded from this workflow.
See third-party notices.
Native plugins are 2.0.0-beta.1. Validation evidence covers planning, bounded dispatch, recovery, file guards, adapter registration and packaging. Live end-to-end execution inside the four client apps is not yet certified. An installable plugin cannot add native spawning tools to a host surface that does not expose them.
npm ci --ignore-scripts
npm test
npm run check
npm run check:packageThe Python eng 1.1 runner is retained for existing users and historical evidence.
It is not needed by any native plugin. Migration
explains how to remove old prompt-routing hooks so they do not intercept native work.
Legacy CLI guide documents that optional runner.