Most tools for building software with AI keep the process we already have — the Scrum sprint, with its stories, gates, and fixed roles — and accelerate it. These two Claude Code skill suites take opposite answers to a sharper question: once the people in the loop are agents, is the sprint worth keeping at all?
| Suite | Premise | In one line |
|---|---|---|
delivery-team |
AI acceleration of a human sprint | 13 sub-agents acting as a complete product-delivery team, orchestrated through an 11-skill sprint lifecycle. |
flow |
AI-first from the ground up | The spec is the source of truth; code is regenerated output. One wide probe generation of parallel variants, multi-objective Pareto evaluation, an evidence-gated ship instead of a calendar — with weight-classed cost envelopes and mandatory token budgets. |
delivery-team keeps the shape of a Scrum sprint — stories, gates, retros, fixed roles — and makes every seat faster. flow discards that scaffolding and rebuilds the pipeline around what LLM agents are uniquely good at (parallel reading, structured scoring, deterministic projection) and bad at (parallel writes, free-form coordination, drifting prose).
delivery-team-vs-flow.md is the full side-by-side comparison and the best place to start if you're deciding which to use.
flow/ # AI-first delivery suite
commands/ # slash-command skills (flow-init, flow-spec, flow-generate, …)
agents/ # function-shaped sub-agents (orchestrator, evaluator, generator, …)
context/ # shared reference docs the skills load
README.md # suite overview
USAGE.md # practitioner's guide
delivery-team/ # Sprint-accelerator delivery suite
commands/ # sprint lifecycle skills (dt-start, dt-run, dt-close, …)
agents/ # role-shaped sub-agents (scrum-master, devs, QA, PMs, …)
context/ # pipeline stages, schemas, definition of done, …
README.md # suite overview
shared/
context/ # context files both suites depend on (vault-access, spec-writing-guide)
archive/
flow/ # retired flow skills, kept for the record (see archive/flow/README.md)
install.sh # installs the suites into ~/.claude/
Requires Claude Code.
git clone https://github.com/shinytoyrobots/agentic-delivery-suites.git
cd agentic-delivery-suites
./install.sh # symlinks the suites into ~/.claude/ (edits propagate live)Options:
./install.sh --copy # copy instead of symlink
./install.sh --suite flow # install just one suite
./install.sh --suite delivery-teamThe installer places commands in ~/.claude/commands/, context in ~/.claude/commands/context/, and agents in both ~/.claude/agents/ (native subagents) and ~/.claude/commands/agents/. Restart Claude Code afterward. Invoke skills as slash commands, e.g. /flow-init or /dt-start.
The suites run standalone on the local filesystem. Several agents can optionally use MCP servers if you have them connected — referenced in agent frontmatter and prose by generic names you can map to your own servers:
| Generic name | What it's used for |
|---|---|
linear |
Issue / cycle tracking (Scrum Master sync) |
notion, slack |
Status, comms, and coordination surfaces |
design-system |
Design-token / component lookup (frontend & design agents) |
notes |
A notes / knowledge-base store for research and bookmarks |
figma, github, hubspot |
Design, repo, and CRM context |
None are required. Where an integration is absent, the agents fall back to local files (see shared/context/vault-access.md). Durable artifacts are written to a local docs/ directory by default — configurable to any folder you like.
These are experimental research artifacts, shared as-is. They encode a specific point of view about agentic delivery; treat them as a starting point to adapt, not a finished product. Issues and forks welcome.
MIT — see LICENSE.