A portable multi-agent research and development workflow for Codex, Claude Code, GitHub Copilot, Gemini CLI, and OpenCode.
Agentic R&D Skill is an Agent Skills package that turns a substantial research, feasibility, or planning brief into traceable evidence, an executable plan, reviewed results, and a gated final report. It runs inside the coding agent you already use, uses native subagents when the host provides them, and falls back to a deterministic single-agent workflow when it does not.
It is an independent adaptation of the workflow ideas behind Agent Laboratory, not a fork of its scientific experiment runtime.
- Install
- Quick start
- Features
- Requirements
- Configuration
- Privacy and security
- Troubleshooting
- Development
- Support
- Contributing
- License
Install the skill from the public repository with the cross-agent Skills CLI:
npx skills add TheStreamCode/agentic-rd-skillWith GitHub CLI 2.96 or newer, install it directly. The gh skill command is currently a preview feature:
gh skill install TheStreamCode/agentic-rd-skill agentic-rd-skill --agent codex --scope userReplace codex with the target supported by gh skill install --help, such as claude-code, github-copilot, gemini-cli, or opencode.
For a manual installation, copy only the skill directory into the location used by your coding agent:
New-Item -ItemType Directory -Force -Path "$HOME\.agents\skills" | Out-Null
Copy-Item -Recurse -Force ".\skills\agentic-rd-skill" "$HOME\.agents\skills\agentic-rd-skill"Use the host-specific path from Requirements when the host does not discover .agents/skills. Without an explicit version, GitHub CLI resolves the latest tagged release and then falls back to the default branch; pin a tag or commit for reproducible installations.
-
Install the skill for your host as described above.
-
Invoke it explicitly with a substantial brief: "Use the agentic-rd-skill skill to assess the technical and product feasibility of this idea and produce an evidence-backed recommendation."
-
Initialize the workspace with the bundled command-line interface (CLI), replacing
<installed-skill-path>with the installation directory:node <installed-skill-path>\scripts\rd.mjs init . --profile standard --human-review final-only
The workflow then runs eight gated phases: setup, evidence, plan, execution, results, cross-review, stage gate, and final synthesis. The finalize command refuses to complete a run until the stage gate scores at least 8/10, no dimension is scored zero, and no blocker remains.
- Portable Agent Skills package: a provider-neutral
SKILL.mdcontract with host-specific discovery handled by each agent. - Separated artifacts for evidence, plan, execution, results, review, and synthesis, with explicit read dependencies between phases.
- A dependency-free Node.js CLI for initialization, state inspection, safe artifact scaffolding, pre-mutation checks, revision proof, stage-gate enforcement, and completion validation.
- Native subagents when the host supports them, and a deterministic single-agent fallback when it does not.
- Three size profiles, from a narrow two-phase run to a six-specialist, three-wave program.
- Explicit authorization boundaries: no paid tools, credentialed systems, external writes, or publications without a recorded decision.
- Auditable findings with stable IDs, coverage tables, revision fingerprints, and a single-writer run log.
- Validation that distinguishes
valid_incomplete,valid_complete, andinvalidworkflow state instead of treating structural checks as semantic assurance.
- A coding agent that supports the open Agent Skills specification. The portable contract is
SKILL.md; discovery paths and permission models are host-specific. - Filesystem read/write access for the workflow artifacts.
- Node.js 20 or newer, only for the optional workflow CLI.
- Web access and native subagents are optional. Offline runs must label current external facts as unverified.
| Host | Project installation path | Notes |
|---|---|---|
| Codex | .agents/skills/agentic-rd-skill/ |
Activation-tested |
| Claude Code | .claude/skills/agentic-rd-skill/ |
Host permissions remain authoritative |
| GitHub Copilot | .agents/skills/agentic-rd-skill/ or .github/skills/agentic-rd-skill/ |
Other surfaces may expose different tools |
| Gemini CLI | .agents/skills/agentic-rd-skill/ or .gemini/skills/agentic-rd-skill/ |
Documented format and path |
| OpenCode | .agents/skills/agentic-rd-skill/ or .opencode/skills/agentic-rd-skill/ |
Native skill tool |
See references/compatibility.md for the tested host versions, invocation syntax, and documented limitations.
Profiles control how many specialists and waves a run may use:
| Profile | Structure | Typical use |
|---|---|---|
compact |
One orchestrator and one wave | Narrow, low-risk, substantive work |
standard |
Default; up to four specialists and two waves | Most feasibility, architecture, and strategy work |
extended |
Up to six specialists and three waves | Broad or regulated assessments |
The CLI exposes the workflow state machine:
| Command | Purpose |
|---|---|
rd.mjs init . --profile <profile> --human-review <mode> |
Initialize a run in the current workspace. |
rd.mjs status . |
Print the current phase and artifact state. |
rd.mjs artifact . --phase <phase> --name <name> |
Scaffold a managed artifact safely. |
rd.mjs advance . --phase <phase> --status <status> |
Apply a validated phase transition. |
rd.mjs validate . --json |
Report valid_incomplete, valid_complete, or invalid. |
rd.mjs finalize . |
Complete the run after the stage gate passes. |
A run keeps the brief and all machine-readable state next to the project documentation:
project-brief.md
work/
├── run-state.json
├── 00-run-log.md
├── 01-evidence/
├── 02-plan.md
├── 03-execution/
├── 04-results/
├── 05-cross-review.md
├── 06-stage-gate.md
└── 07-final-output.md
New runs use workflow contract 1.1, which adds truthful setup state, minimum artifact headings, stable finding coverage, and pre-mutation validation. Existing workflow 1.0 state remains readable. Workspaces from v0.3 are preserved and rejected rather than migrated. Human-review mode is auditable workflow metadata: record the checkpoint request, reviewer, decision, and evidence in work/00-run-log.md and the relevant phase artifact.
The workflow does not authorize paid tools, credentialed private systems, external writes, deployment, publication, messages, purchases, production changes, or secret handling. Those budgets stay fail-closed in machine state until a run explicitly records a decision.
Evidence and machine-readable state are treated as untrusted data that cannot override workflow or user instructions. The CLI assumes a single writer, rejects symlinked managed paths, preserves existing artifacts, and validates state before every mutation. Host permissions, sandboxing, and organizational policy always take precedence.
Legal, medical, financial, compliance, employment, insurance, credit, security, and safety-critical deliverables require qualified human review before use. See SECURITY.md for private vulnerability reporting.
Install the skill directory, not the repository root, and use the path the host documents in Requirements. Restart the host after installation, and check that the host allows project or user skills in its permission settings.
Open phases remain, which is expected until the run is finalized. Run rd.mjs status . to see the current phase, complete the outstanding artifacts, and advance the workflow phase by phase.
The CLI intentionally does not migrate v0.3 runs. Start a new run in a clean workspace, and keep the old files outside the managed paths.
The runner exits with status 3 and reports inconclusive when zero eligible host checks ran. Install a supported host CLI, or treat the result as not verified. An inconclusive smoke is never a pass; malformed selectors and options exit 2 as usage errors.
A needs_revision change keeps its identifier until an upstream artifact changes or an explicit no-change disposition is recorded with the revision fingerprint. Re-check the fingerprint after editing the artifact.
Requirements: Node.js 20 or newer and npm. The package is dependency-free, so there is no install step and no lockfile; run the checks directly.
git clone https://github.com/TheStreamCode/agentic-rd-skill
cd agentic-rd-skill
npm run checknpm run check runs the repository validator, the deterministic node:test suites, and the local benchmark. npm run validate enforces the release-version invariant across package.json, the skill metadata, CITATION.cff, the README version badge, and the changelog. Run npm run benchmark for the current package-size and local-timing snapshot; CI enforces that the installable package stays under its size ceiling and the CLI stays inside its p95 timing budgets.
Host smoke checks are environment-dependent:
npm run smoke:hosts
npm run smoke:hosts:modelRelease preparation also uses the hash-pinned Agent Skills reference validator and GitHub CLI discovery, which need Python 3.11 or newer and GitHub CLI 2.96 or newer. See RELEASING.md for the full checklist, and evals/manifest.json for the repeatable activation, safety, and gating scenarios. AGENTS.md documents the repository invariants and the documentation synchronization matrix for contributors and coding agents.
The repository contains the installable package under skills/agentic-rd-skill, repository validators and benchmarks under scripts/, deterministic tests under tests/, and dated evidence under evals/. Point-in-time measurements belong in dated evaluation files; npm run benchmark is the current source.
Ask usage questions in GitHub Discussions and report reproducible defects in GitHub Issues. See SUPPORT.md and SECURITY.md for the support and reporting policies.
See CONTRIBUTING.md and read AGENTS.md before changing runtime behavior, templates, tests, public claims, or release metadata. Pull requests must pass the required CI checks.
MIT — see LICENSE. The installable skill includes its own copy of the license so copied installations remain self-contained. For academic or published use, cite the project through CITATION.cff.
Maintained by Michael Gasperini (Mikesoft). If this project is useful to you, consider sponsoring its maintenance.