agentic-workflow

A public orchestration pack for building cleaner, safer, more verifiable multi-agent workflows.

This repository is a generic, reusable Agent OS / workflow pack for structured multi-agent systems. It packages the parts that usually get hand-waved away in ad hoc setups:

• shared contracts
• role templates
• starter bundles
• escalation patterns
• verification rules
• permission profiles
• trust-boundary rules
• handoff snippets
• specialist role overlays

The goal is simple: make multi-agent work less vague, less fragile, and less fake-done.

---

What this repo is

This pack is for operators who want a stronger baseline for:

• approval-first execution
• bounded delegation
• critique and verification lanes
• cleaner handoffs between roles
• more disciplined completion rules

Instead of throwing a few agent prompts together and hoping they coordinate well, this repo gives you a more structured starting point.

Important

This is a public baseline pack, not a private deployment dump. It stays generic on purpose so you can fork it, adapt it, and layer your own local rules on top.

---

What makes it different

Area	Default stance
New work	approval-first
Delegation	bounded, single-purpose, explicit stop conditions
Completion	worker completion is not final completion
Quality control	critique lane + verification lane
Recovery	explicit retry / diagnose / repair / verify loop
Permissions	role-scoped access instead of broad default power
Trust model	direct operator intent beats retrieved content
Tone	neutral, operational, reusable

In other words: this pack treats planning, execution, critique, and verification as different jobs.

---

Core role model

Role	Job
primary-agent	front door, coordinator, release owner
planner	breaks work into bounded steps
researcher	gathers evidence, sources, uncertainty
executor	performs scoped action and repair work
critic	finds defects and writes concrete fix tickets
verifier	checks acceptance criteria and evidence

Why split critic and verifier?

Because these are not the same job.

• critic asks: what is still wrong, risky, weak, or incomplete?
• verifier asks: did we actually meet the acceptance criteria with evidence?

That separation is one of the main things that keeps a workflow from drifting into fake confidence.

---

Specialist role library

Beyond the core workflow spine, this repo now includes a broader copy-paste role library for operators who want more than a minimal setup.

Category	Included specialists
Engineering	frontend-developer, backend-architect, devops-automator, security-engineer, software-architect, technical-writer, rapid-prototyper, sre
Product	product-manager, sprint-prioritizer, feedback-synthesizer, trend-researcher, behavioral-nudge-engine
Testing	reality-checker, accessibility-auditor
Support	analytics-reporter, infrastructure-maintainer, executive-summary-generator, support-responder, legal-compliance-checker, finance-tracker
Growth	paid-social-strategist, content-creator, outbound-strategist
Design	ui-designer, ux-researcher, ux-architect, brand-guardian, visual-storyteller, whimsy-injector
Paid Media	ppc-strategist, tracking-measurement-specialist, ad-creative-strategist, search-query-analyst, paid-media-auditor
Sales	discovery-coach, deal-strategist, pipeline-analyst, proposal-strategist, sales-engineer, account-strategist
Project Management	project-shepherd, studio-producer, experiment-tracker, jira-workflow-steward, project-manager-senior

These are optional overlays. The repo still expects the core orchestration spine to stay intact.

At a glance

• core workflow roles: 6
• support / vertical baseline roles: 2
• specialist overlay roles: 46
• total role templates in repo: 54

Starter bundles

You do not need every role for every workflow.

This repo includes pre-shaped starter bundles so an operator can begin small and scale only when needed.

Bundle	Best for	Includes
MINIMAL	simple direct work	primary-agent, executor, verifier
STANDARD	repeated multi-step work	primary-agent, planner, executor, verifier
RESEARCH-HEAVY	evidence-first work	primary-agent, planner, researcher, executor, verifier
QUALITY-CRITICAL	failure-sensitive work	primary-agent, planner, executor, critic, verifier
MARKETING-STARTER	messaging / campaign workflows	primary-agent, planner, executor, verifier, marketing-strategist

If you're unsure where to begin, the intended path is:

1. use COPY-THIS-PROMPT.md
2. let the primary agent choose the smallest suitable starter bundle
3. approve before anything is applied

---

Repository layout

Expand file tree

agent-os-pack/
├── README.md
├── START-HERE.md
├── COPY-THIS-PROMPT.md
├── INTERNAL-BUNDLE-PICKER.md
├── shared/
│   ├── APPROVAL_BOUNDARIES.md
│   ├── ESCALATE.md
│   ├── MEMORY_POLICY.md
│   ├── ORCHESTRATION.md
│   ├── PERMISSION_PROFILES.md
│   ├── PLACEHOLDERS.md
│   ├── QUESTION_ROUTING.md
│   ├── REPORTING_CONTRACT.md
│   ├── TRUST_BOUNDARIES.md
│   └── VOICE_SYSTEM.md
├── packs/
│   ├── core/
│   │   ├── primary-agent/
│   │   ├── planner/
│   │   ├── researcher/
│   │   ├── executor/
│   │   ├── critic/
│   │   └── verifier/
│   ├── support/
│   ├── vertical/
│   └── specialists/
│       ├── design/
│       ├── engineering/
│       ├── growth/
│       ├── paid-media/
│       ├── product/
│       ├── project-management/
│       ├── sales/
│       ├── support/
│       └── testing/
├── starter-bundles/
│   ├── MINIMAL.md
│   ├── STANDARD.md
│   ├── RESEARCH-HEAVY.md
│   ├── QUALITY-CRITICAL.md
│   └── MARKETING-STARTER.md
├── snippets/
└── examples/

Folder guide

• shared/ → deployment-wide rules, permissions, trust boundaries, and contracts
• packs/core/ → main role templates
• packs/support/ → helper roles such as intake routing
• packs/vertical/ → domain-specific examples
• packs/specialists/ → broader copy-paste specialist library
• starter-bundles/ → pre-shaped bundle recommendations
• snippets/ → fast copy/paste blocks for handoffs and escalation
• examples/ → composition examples and customization checklists

---

Quick start

1) Start small

Read:

• START-HERE.md
• COPY-THIS-PROMPT.md

2) Let the primary agent choose a bundle

Use the prompt in COPY-THIS-PROMPT.md to have the system recommend the smallest suitable starter bundle.

3) Approve before applying

The baseline posture here is approval-first for new work.

4) Customize locally

Before production use, tighten:

• approval boundaries
• tool permissions
• trust-boundary rules for untrusted content
• data-handling rules
• subagent limits
• domain guardrails

---

Philosophy

This pack is built around a few simple operating beliefs:

• planning is not completion
• execution is not verification
• worker-reported Completed is not objective-level DONE
• approval should happen before new work starts
• delegation should be explicit, narrow, and reviewable
• completion claims should be backed by evidence
• retrieved content should never outrank direct operator intent

If your current workflow feels like a pile of good intentions, this repo is meant to give it a stronger spine.

The specialist library exists so operators can start with a clean core and then pull in deeper domain roles only when they are actually needed.

---

Best use cases

This repo is a good fit when you want to build:

• a cleaner personal Agent OS baseline
• a reusable multi-agent prompt pack
• a more disciplined subagent workflow
• a verification-aware execution loop
• a safer orchestration baseline for real work

It is not trying to be:

• a private dump of one operator's local system
• a full beginner setup tutorial for any one host platform
• a giant one-size-fits-all enterprise framework

---

Customization checklist

Before treating this pack as deployment-ready:

• replace required placeholders in shared/PLACEHOLDERS.md
• tighten shared/APPROVAL_BOUNDARIES.md
• set subagent rules in shared/ORCHESTRATION.md
• decide which roles are active and which are removed
• add vertical guardrails if you enable domain-specific roles
• test blocked, low-risk, and multi-step delegated scenarios

For the full checklist, see:

• examples/CUSTOMIZATION_CHECKLIST.md

---

Suggested first files to open

If you want the cleanest path through the repo, go in this order:

1. START-HERE.md
2. COPY-THIS-PROMPT.md
3. INTERNAL-BUNDLE-PICKER.md
4. shared/ORCHESTRATION.md
5. starter-bundles/STANDARD.md

---

Why this repo exists

Because a lot of multi-agent setups fail in predictable ways:

• too many roles too early
• unclear approval boundaries
• vague subagent handoffs
• confident reporting without verification
• “completed” work that is not actually done

This pack is a public attempt to make those failure modes harder to ignore.

---

Third-party adaptation note

Some specialist roles in packs/specialists/ are adapted from the MIT-licensed msitarzewski/agency-agents project and normalized into this repository's generic, third-person, approval-first format. See THIRD_PARTY_NOTICES.md.