b2b-agents

B2B Agents

A white-label, platform-neutral multi-agent system. 56 specialized agents covering engineering, product, marketing, design, operations, and testing — plus a company builder that configures the whole system for your organization in one session, and a compile pipeline that targets the platform you actually run on.

Author your agents once. Ship them to Claude Code, the OpenAI Agents SDK, or a Custom GPT — the pipeline handles the differences and tells you exactly what changes per platform.

What’s new in this version

Earlier versions were Claude Code only. This version separates the agent content (the portable persona, knowledge, and examples) from the platform wiring (model names, tool bindings, orchestration mechanism). A small pipeline compiles the neutral source to each target:

.md agents ──(extract)──► neutral/*.agent.yaml ──(adapter)──► Claude Code  (.md)
                                                ├─────────────► OpenAI Agents SDK (Python)
                                                └─────────────► Custom GPT (instruction blobs)

• One source, many platforms. The same agent library runs on full multi-agent frameworks or a single no-code GPT.
• Honest degradation. Limited platforms (e.g. Custom GPTs, which can’t orchestrate) get a build report listing exactly what was flattened, dropped, or needs manual setup — nothing fails silently.
• company-builder picks the target for you. It asks where you’ll run the agents during intake and compiles to the matching platform(s) automatically.

See docs/neutral-agent-schema.md for the full design and scripts/README.md for the pipeline.

Quick Start

git clone https://github.com/alyssonfranklin/b2b-agents.git
cd b2b-agents

Then, in Claude Code:

Use the company-builder agent

It asks ~9 questions (including which platform you run on), generates your company folder, recommends which agents fit your workflow, rewrites CLAUDE.md with your context, compiles to your target platform(s), and produces an install path.

Manual install for Claude Code (skip the builder)

cp -r b2b-agents/*.md ~/.claude/agents/
for d in engineering product marketing design project-management studio-operations testing bonus; do
  cp -r "b2b-agents/$d" ~/.claude/agents/
done

Or copy only the agents relevant to your workflow — fewer agents means faster routing and less decision overhead.

Multi-Platform Pipeline

The agent .md files are the source of truth. To run on another platform, compile them.

Target	Capability	Output
claude	Native subagents, full fidelity	Claude Code .md agents
sdk	Code-orchestrated multi-agent	OpenAI Agents SDK package (also fits LangGraph / CrewAI)
gpt	Single-agent, instructions-only	Custom GPT / Gemini Gem instruction blobs

pip install pyyaml --break-system-packages

python3 scripts/extract_neutral.py --src .       --out neutral      # agents → neutral
python3 scripts/adapter_claude.py  --src neutral --out dist/claude  # → Claude Code
python3 scripts/adapter_sdk.py     --src neutral --out dist/sdk     # → OpenAI Agents SDK
python3 scripts/adapter_gpt.py     --src neutral --out dist/gpt     # → Custom GPT

What survives the trip:

Neutral concept	claude	sdk	gpt
Reasoning tier (fast/balanced/deep)	haiku / sonnet / opus(+thinking)	model tier + reasoning_effort	manual model choice (warns on deep)
File/web/shell capabilities	native tools	runtime function tools + hosted tools	Code Interpreter / Knowledge / Web Browsing (warns where unsupported)
Orchestration	Task delegation	handoffs	flattened into one agent + warning
Structured input (input_schema)	passthrough	documented	“Expected input” section

The claude round-trip is verified lossless (neutral → claude → neutral is byte-identical). neutral/ and dist/ are generated and git-ignored — rebuild them, don’t edit by hand.

Architecture

Repository structure

b2b-agents/
├── company-builder.md           # Run this first — intake, customize, compile
├── b2b-orchestrator.md          # Multi-domain task coordinator
├── company/                     # Template agents (filled in by the builder)
├── space-industries/            # Reference example (DXP consulting firm)
├── engineering/                 # 16 agents
├── product/                     # 7 agents
├── marketing/                   # 9 agents
├── design/                      # 5 agents
├── project-management/          # 3 agents
├── studio-operations/           # 7 agents
├── testing/                     # 5 agents
├── bonus/                       # 2 agents
├── scripts/                     # Compile pipeline (extractor + 3 adapters)
├── docs/
│   ├── neutral-agent-schema.md  # The multi-platform schema + adapter spec
│   ├── architecture.md          # How the agent system works
│   ├── writing-agents.md        # How to author effective agents
│   ├── eval/                    # Test prompts per agent category
│   └── compliance/              # ISO/ISMS documentation (not agents)
├── neutral/                     # Generated: neutral agent YAML (git-ignored)
└── dist/                        # Generated: per-platform builds (git-ignored)

space-industries/ shows what a completed company setup looks like for a DXP consulting firm. Use it as a reference when customizing your own.

Reasoning tiers

Each agent declares a reasoning tier rather than a hardcoded model, so the same agent maps to the right model on every platform. On the claude target the tiers resolve as:

Tier	Claude model	Used for
deep	claude-opus-4-6 (+ thinking)	Security review, legal, compliance, backend architecture, AI engineering, pricing strategy
balanced	claude-sonnet-4-6	Most engineering, product, design, marketing, testing agents
fast	claude-haiku-4-5-20251001	Social content, support responses, coordination, jokes

On sdk the deep tier requests reasoning_effort="high"; on gpt it prompts you to pick a reasoning model by hand.

Extended thinking (deep tier)

These agents reason with a 10k-token thinking budget on Claude: enterprise-security-reviewer, backend-architect, ai-engineer, data-privacy-engineer, legal-advisor, compliance-automation-specialist, pricing-packaging-strategist.

Orchestrator

b2b-orchestrator.md coordinates tasks that span more than one department. Its prompt describes delegation generically (“delegate to a specialist”), so it compiles correctly to each platform — Task on Claude, handoffs on the SDK, and an inlined single agent on Custom GPT.

"We're launching enterprise SSO next week — handle everything"
→ decomposes → delegates to security-reviewer + devops-automator + content-creator +
  project-shipper in parallel → synthesizes into a launch brief

Company Builder

company-builder.md runs at fork time (or any time you reconfigure):

1. Intake — 9 questions: name, type, domain, tech stack, business model, deal size, top use cases, and the platform you’ll run on.
2. Recommendation — maps your answers to the relevant subset of the agents.
3. Generation — creates [your-company]/ with the 4 business agents customized to your context, and rewrites CLAUDE.md.
4. Compile — builds for your chosen platform(s) and surfaces any degradation report.
5. Install — for Claude, an install.sh; for SDK/GPT, the compiled package or blobs.

Agent Catalog

Orchestrator

• b2b-orchestrator — Decomposes complex multi-domain tasks, coordinates specialists, synthesizes results

Company (generated by builder, see space-industries/ for a reference example)

• proposal-writer — Proposals, SOWs, retainer agreements in your company’s format and voice
• discovery-facilitator — Discovery workshops, stakeholder interviews, findings reports
• rfp-powerhouse (deep) — Full RFP assembly, compliance matrices, win strategy
• client-presentation-builder — Executive decks, status updates, proposal presentations

Engineering (16 agents)

• accessibility-expert — WCAG 2.1, Section 508, LGPD accessibility compliance
• ai-engineer (deep) — LLM integration, ML pipelines, RAG systems
• ai-ethics-governance-specialist — AI bias detection, algorithmic transparency
• api-integration-specialist — REST/GraphQL design, SDK development, developer experience
• backend-architect (deep) — Scalable B2B APIs, multi-tenant systems, SSO
• data-privacy-engineer (deep) — GDPR, CCPA, LGPD, privacy-by-design
• database-performance-optimizer — Multi-tenant DB optimization, query tuning
• devops-automator — CI/CD pipelines, cloud infrastructure, IaC
• enterprise-integration-architect — Salesforce, SAP, Microsoft 365 integrations
• enterprise-security-reviewer (deep) — SOC2, GDPR, ISO 27001 compliance
• frontend-developer — React/Vue/Angular enterprise UIs
• mobile-app-builder — B2B React Native applications
• monitoring-observability-specialist — APM, SLA monitoring, distributed tracing
• rapid-prototyper — B2B MVPs and enterprise feature POCs
• technical-sales-engineer — Demos, RFP technical responses, solution architecture
• test-writer-fixer — Enterprise test suites, failure diagnosis, test repair

Product (7 agents)

• b2b-onboarding-optimizer — Activation flows, time-to-value, enterprise setup
• enterprise-onboarding-strategist — Large-scale enterprise rollouts, change management
• feedback-synthesizer — Customer feedback analysis, pattern extraction
• pricing-packaging-strategist (deep) — Value-based pricing, enterprise contracts
• product-sales-specialist — Product design for sales, RFP creative responses
• sprint-prioritizer — Backlog prioritization, feature trade-offs, sprint planning
• trend-researcher — B2B market opportunities, competitive intelligence

Marketing (9 agents)

• app-store-optimizer (fast) — B2B SaaS app store presence
• content-creator — Blog posts, thought leadership, multi-format content
• enterprise-content-strategist — B2B content strategy, SEO, buyer journey
• growth-hacker — Viral loops, PLG, enterprise acquisition
• linkedin-content-curator (fast) — LinkedIn posts, articles, thought leadership
• partnership-channel-manager — B2B partnerships, channel programs
• reddit-community-builder (fast) — Authentic Reddit community engagement
• tiktok-strategist (fast) — B2B TikTok content strategy
• twitter-engager (fast) — Twitter/X B2B engagement and content

Design (5 agents)

• brand-guardian — B2B brand identity consistency
• ui-designer — Enterprise UI components, design systems
• ux-researcher — Enterprise user research, usability testing
• visual-storyteller — Data visualization, pitch decks
• whimsy-injector (fast) — Professional delight and micro-interactions

Project Management (3 agents)

• experiment-tracker — A/B testing, feature flag management
• project-shipper — B2B feature launches, go-to-market coordination
• studio-producer — Cross-team coordination, resource management

Studio Operations (7 agents)

• analytics-reporter — B2B usage data, business metrics dashboards
• compliance-automation-specialist (deep) — Automated SOC2, ISO 27001
• customer-success-manager — Enterprise account health, churn prevention, expansion
• finance-tracker — MRR, unit economics, B2B financial modeling
• infrastructure-maintainer — Cloud cost, capacity planning, infrastructure scaling
• legal-advisor (deep) — Enterprise contracts, RFP legal, procurement compliance
• support-responder (fast) — B2B customer support, issue resolution

Testing (5 agents)

• api-tester — B2B API load testing, enterprise reliability
• performance-benchmarker — Application performance, enterprise load testing
• test-results-analyzer — Test failure pattern analysis
• tool-evaluator — B2B development tool assessment and selection
• workflow-optimizer — Human-agent collaboration, process automation

Bonus (2 agents)

• studio-coach (fast) — Team motivation, multi-agent coordination support
• joker (fast) — Tech humor for stressed teams

Authoring Agents

Agents are authored as Claude .md files (the source of truth); the pipeline lifts them into neutral form for compiling to other platforms. Keep agent bodies platform-neutral — no tool names, no model names, no runtime references. Refer to delegation generically (“delegate to a security specialist”), and the adapters bind it per platform.

---
name: your-agent-name
description: Use this agent when [scenario]. Examples:

<example>
Context: [situation]
user: "[request]"
assistant: "[what you'll do — no self-reference]"
<commentary>
[why this example matters]
</commentary>
</example>
color: blue
model: claude-sonnet-4-6
tools: Read, Write, Bash, Grep
---

You are a [role] specializing in [domain]...

For deep-reasoning agents add extended thinking; for agents invoked with structured inputs add an input_schema:

model: claude-opus-4-6
thinking:
  type: enabled
  budget_tokens: 10000
input_schema:
  type: object
  properties:
    required_field:
      type: string
      description: "What this field means and example values"
  required:
    - required_field

After editing any agent, re-run the pipeline so the per-platform builds stay in sync.

Documentation

Doc	What it covers
docs/neutral-agent-schema.md	The multi-platform schema, the three target buckets, and adapter mappings
scripts/README.md	The compile pipeline: extractor + adapters, usage, per-target output
docs/architecture.md	How the agent system works: routing, frontmatter, orchestration, input_schema, white-label layers
docs/writing-agents.md	Writing effective agents: description as routing logic, examples, model/tool selection, prompt design
docs/agent-intelligence-architecture.md	Blueprint for evolving the system: Memory, Hierarchical Specialization, Quality Gates, Context Amplifiers
docs/eval/README.md	How to run agent evaluations

Notes

• neutral/ and dist/ are generated by the pipeline and git-ignored — rebuild, don’t commit.
• docs/compliance/ contains ISO/ISMS documentation — not agents.
• marketing/instagram-curator.md is deprecated; use marketing/linkedin-content-curator.md.
• base22/ is deprecated; see space-industries/ for the reference example.
• space-industries/ is a generated example — a DXP consulting firm configured via company-builder.