Agenticflow Skills

AgenticFlow Skills

AgenticFlow is a platform for building AI-powered automation workflows, intelligent agents, and workforce systems.

This skill is loaded by AI agents operating AgenticFlow on a user's behalf — Ishi (AgenticFlow's first-party desktop agent), Claude Code, OpenAI Codex, Cursor, Gemini CLI, or any compatible host. The AgenticFlow CLI (af) is the shared contract every host uses. See the Ecosystem overview for how the layers fit together.

First-Time Setup (for AI agents)

Invocation priority — try these in order, use the first that works. The CLI installs as TWO binaries: agenticflow (canonical, 11 chars, collision-safe) and af (2-letter shortcut, sometimes claimed by unrelated tools like Python packages, homebrew formulas, shell aliases).

# 1. Preferred: full canonical binary. Rare to collide.
command -v agenticflow >/dev/null 2>&1 && agenticflow --version

# 2. Fallback: npx with the full package name. Works even with nothing installed.
npx --yes @pixelml/agenticflow-cli --version

# 3. Only use `af` if you've confirmed it's ours (prints a semver, not a Python traceback):
command -v af >/dev/null 2>&1 && af --version

Critical — never use npx af. af is a generic 2-letter name and npm has an unrelated package named af. npx af fetches the wrong package. Always use the full spec: npx --yes @pixelml/agenticflow-cli <subcommand>.

Once you've picked an invocation (call it AF), use it consistently for every command:

AF bootstrap --json     # orient — returns auth, agents, workforces, blueprints, etc.

Always use --json on all commands for structured output you can parse. After bootstrap, extract _links from the JSON output and present web UI URLs to the user.

Decision Policy

Use this to decide WHEN to run each operation:

When Things Go Wrong

Quick Navigation

Workflow

Workflows are linear automation pipelines composed of sequential nodes. Each node performs a specific action.

Node Types Overview

Note: Workflows in AgenticFlow are linear and sequential - nodes execute top to bottom with no branching or loops.

Agent

An Agent is an AI entity with specific capabilities, tools, and a defined persona.

To learn about agent configuration, load: reference/agent/overview.md

For CLI-first agent operations (create/get/update/stream), load: reference/agent/cli-mode.md

The composition ladder

AgenticFlow's three primitives — workflow, agent, workforce — are rungs on a complexity ladder. Start at the lowest rung that solves the user's problem. Each rung composes from the rungs below.

Routing rule for AI operators: if the user's need is deterministic and stepwise, pick workflow. If it needs tool-picking flexibility, pick agent. If it needs explicit multi-agent coordination, pick workforce.

af blueprints list [--kind <k>] [--complexity <n>] --json returns every shipped blueprint with its rung. af bootstrap --json > blueprints[] surfaces the same data inline.

Blueprints

A blueprint is a pre-made starter pattern. As of CLI v1.10.0 (2026-04-14) blueprints span the full ladder (rungs 0-3, 6):

To see blueprint details, load: reference/blueprints.md

🕰️ Heads-up on the "pack" concept: AgenticFlow used to ship a separate af pack * surface for pre-built business kits. As of CLI v1.7.0 (2026-04-14), the three original packs are all available as blueprints, and af pack * is deprecated with a 2026-10-14 sunset. If you see a user reach for af pack install, redirect them to af workforce init --blueprint <id> (or af agent init --blueprint <id> for Tier 1) — same content, one verb, one catalog.

Available Blueprints (20 — across rungs 0-3 and 6)

Rungs 0-2: Workflow blueprints — deterministic chains. Need one LLM-provider connection (auto-discovered):

Deploy: af workflow init --blueprint <id>.

Rung 3: Agent blueprints — single agent with built-in plugins (work in any workspace):

Deploy: af agent init --blueprint <id>.

Rung 6: Workforce blueprints — multi-agent DAGs.

Batteries-included (plugins pre-attached — work end-to-end with zero setup):

Deploy: af workforce init --blueprint <id>.

Rung 6: Vertical workforce blueprints — generic agents, attach your own MCP tools after deploy:

Quick Start — Rung 0 (simplest workflow)

# Dry-run — see the workflow create payload
af workflow init --blueprint llm-hello --dry-run --json

# Deploy
af workflow init --blueprint llm-hello --json
# → returns { workflow_id }

# Run (pass the trigger's named input)
af workflow run --workflow-id <id> --input '{"question":"What is a unicorn?"}' --json
# → returns { id: run_id }
af workflow run-status --workflow-run-id <run_id> --json
# → poll until status=success; `.output.content` is the LLM's answer

Quick Start — Rung 3 (agent + plugins, zero setup)

# Dry-run — prints the agent + plugin config
af agent init --blueprint research-assistant --dry-run --json

# Deploy — creates ONE agent with all plugins attached (~4 plugins for research-assistant)
af agent init --blueprint research-assistant --json
# → returns { agent_id, plugins: [...], _links.agent }

# Smoke-test — runs the agent with a real query that exercises web_search
af agent run --agent-id <id> --message "Latest news about X?" --json

Quick Start — Rung 6 (multi-agent workforce)

# Preview — shows which agents will be created and estimated node/edge counts
af workforce init --blueprint tutor --name "My Tutoring Team" --dry-run --json

# Deploy — creates the workforce, all required agents, and the wired DAG in one call
af workforce init --blueprint tutor --name "My Tutoring Team" --json

# Include optional slots (e.g. the fifth agent in tutor/freelancer/amazon-seller)
af workforce init --blueprint tutor --name "My Tutoring Team" --include-optional-slots --json

# Override the default model for all created agents
af workforce init --blueprint tutor --name "My Tutoring Team" --model agenticflow/gemma-4-31b-it --json

# After the call you get {workforce_id, agents:[{slot_role, agent_id, title}], ...}
# Publish a public URL so your teammates can run it without platform auth:
af workforce publish --workforce-id <id> --json

Blueprint vs Marketplace

Blueprints are offline, versioned, CLI-shipped. The marketplace (af marketplace *, new in v1.8.0) is the live user/admin-curated catalog — three template kinds (agent_template, workflow_template, mas_template) unified under one endpoint. Browse: af marketplace list --limit 50 --json. Clone: af marketplace try --id <item> --dry-run --json. Full comparison: af playbook marketplace-vs-blueprint or load reference/marketplace.md.

Custom business (no blueprint fits)

If the user's domain isn't in the 8-blueprint list (examples: coaching, consulting, legal, real-estate, agency, SaaS support, vertical-specific ops), blueprints aren't the right shortcut. Load one of the narrow skills:

• Single chat endpoint / one assistant / customer-facing bot → skills/agenticflow-agent/SKILL.md
• Multiple agents that hand off but no blueprint matches → skills/agenticflow-workforce/SKILL.md (custom graph via af workforce create + af workforce deploy --body @graph.json)

Workforce

Workforce systems coordinate multiple agents to solve complex tasks collaboratively.

To understand orchestration patterns, load: reference/workforce/overview.md

Common Patterns

• Supervisor - One agent delegates to specialists
• Swarm - Agents self-organize dynamically
• Pipeline - Sequential agent handoffs
• Debate - Agents discuss to reach consensus

Glossary

For terminology and definitions, see reference/glossary.md.

Quality Gate

Before declaring any workflow or agent task done, enforce: reference/quality/acceptance-criteria.md