Skip to main content
Unlisted page
This page is unlisted. Search engines will not index it, and only users having a direct link can access it.

Business Models and Setups: AI Agent Team Tools

· 16 min read

Research date: 2026-02-20

1. dmux (standardagents/dmux) -- Justin Schroeder (@jpschroeder)​

What it is: A dev agent multiplexer. Runs parallel AI coding agents (Claude Code, Codex, OpenCode) in isolated git worktrees via tmux panes. Press n to create a new pane, type a prompt, pick an agent -- dmux handles the worktree, branch, and agent launch. Press m to merge results back. Supports A/B testing (same prompt, two different agents side-by-side).

Open source: Yes. MIT license. npm package: dmux (npm install -g dmux).

Business model: None for dmux itself. Entirely free, no paid tier, no commercial component. dmux is a side project / internal tool that was open-sourced.

How Schroeder actually makes money: He is the creator of FormKit, an open-source form framework for Vue.js. FormKit has a Pro tier:

  • Production (single domain): $149 one-time
  • Multi-domain: pay-as-you-go monthly/annual subscription
  • Enterprise: starting at $1,250 one-time
  • Free during development, paid when deployed to production domains

The standardagents GitHub org appears to be a separate effort from FormKit. dmux was originally under formkit/dmux before moving to standardagents/dmux, suggesting Schroeder may be building a broader agent tooling brand, but as of now there is no monetization of the agent tooling itself.

Technical setup:

  • tmux for terminal multiplexing
  • git worktrees for isolation (each agent gets its own worktree and branch)
  • Hooks system for worktree automation
  • AI-generated branch names and commit messages
  • Multi-project support per session

Sources:

2. Superconductor -- Sergey Karayev (@sergeykarayev)​

What it is: A cloud platform for running teams of AI coding agents. Write informal tickets, spin up many agents per ticket, each gets its own live app preview, one-click PR the best result. Accessible via web, iOS app, Slack, and mobile.

Open source: No. Superconductor is a closed-source commercial product.

Company: Built by Volition, a product studio co-founded by the original Gradescope team (Sergey Karayev, Arjun Singh, Pieter Abbeel). Gradescope was acquired by Turnitin. The co-founders reunited to start Volition.

Funding: Described as "a small but well-funded team." No public fundraising rounds found on Crunchbase. Self-described funding status only. Likely bootstrapped or angel-funded given the team's prior exit from Gradescope/Turnitin.

Pricing: Not publicly disclosed as of February 2026. The platform appears to still be in an early-access/invite phase. Key detail: users "use your Claude Pro/Max Plan, or your own API keys," meaning Superconductor does not bundle LLM costs -- it is a BYOK (bring your own key) model. This suggests the revenue model is a platform fee on top of the user's own API costs.

Infrastructure: Agent sessions run on compute from third-party partners, primarily Modal and Morph Cloud. Sessions live in the cloud, not on individual machines, and anyone on the team can pick up where someone else left off. Agents run in isolated containers with network sandboxing.

Technical setup:

  • Cloud-based sandboxed environments per agent
  • Live app preview per agent
  • Ticket-based interface (web, mobile, Slack)
  • BYOK for LLM costs
  • Agent benchmarking on your codebase

Sources:

3. Terragon -- Sawyer Hood (@sawyerhood)​

What it was: A cloud-based background agent orchestrator. Ran Claude Code (and later Codex, Amp, Gemini) as background agents in isolated cloud containers. Solved the problem of needing to keep your terminal open while Claude Code runs. Web dashboard, CLI, GitHub comments, and mobile interfaces for task management. Each task got a unique branch; agent work was checkpointed and pushed to GitHub with AI-generated commits and PRs.

Open source: The codebase was released as Apache-2.0 upon shutdown at terragon-labs/terragon-oss. Provided "as-is, with no guarantees of maintenance, support, or completeness."

Business model: SaaS subscription. Specific pricing tiers were not preserved in public records, but the service offered subscription plans with credit top-ups for AI usage. Users could also link their own provider accounts.

Status: SHUT DOWN. Announced shutdown January 16, 2026. Service ended February 9, 2026. New subscriptions were stopped immediately. Existing subscriptions canceled at end of billing period. Outstanding credits refunded.

Why it shut down: Not publicly explained in detail. Sawyer Hood's blog post "The Rise of Codex" discusses how quickly the competitive landscape shifted -- Codex reached 28% of agent usage on Terragon within one month of launch, suggesting the market for agent orchestration was becoming commoditized as providers built their own background agent capabilities (Anthropic's Claude Code background tasks, OpenAI Codex, etc.).

Sources:

4. OpenClaw -- Peter Steinberger​

What it is: A free and open-source AI agent framework. Originally called "Clawdbot" (renamed after Anthropic threatened legal action over similarity to "Claude"), then briefly "Moltbot," then OpenClaw. It wraps an LLM in a persistent daemon connected to 12+ messaging platforms (Telegram, WhatsApp, Discord, Slack, etc.) with a heartbeat scheduler, session management, and persistent memory.

Not primarily a coding agent. OpenClaw is a general-purpose personal AI assistant framework. The multi-agent coding orchestration is a community-driven use case built on top of it, where specialized agents (Scout, Optimizer, Implementer) can be coordinated for complex coding tasks.

Open source: Yes. MIT license. 213,000+ GitHub stars (as of February 2026).

Creator: Peter Steinberger, Austrian developer. Previously founded PSPDFKit (PDF management B2B tool, ~70 employees, ~100M EUR valuation before his exit in 2021). As of February 2026, Steinberger has joined OpenAI, with confirmation that OpenClaw will remain open source under OpenAI's stewardship.

Business model:

  • Software is 100% free. No subscription, no premium tier, no paywall.
  • Revenue comes from two planned channels:
    1. OpenClaw Cloud -- managed hosting service, launching later in 2026, starting at $39/month
    2. Enterprise support -- professional services
  • Real user costs are LLM API fees: $5-30/month typical (light users $5-10, regular $15-30, power users $40-100+)
  • Can run entirely free with local models (Ollama + Llama) on Oracle Cloud free tier

Technical setup for multi-agent coding:

  • Each agent gets its own workspace with SOUL.md, AGENTS.md, optional USER.md
  • Dedicated agentDir and session store per agent
  • Agent Teams enable parallel task execution with shared state
  • Inter-agent messaging with @mentions
  • Flexible coordination modes
  • ClawHub skill registry for sharing agent skills

Sources:

5. swarm-code (@999x_ai)​

What it is: Could not find a specific project or repository matching "swarm-code" by "@999x_ai." Multiple searches across GitHub, Twitter/X, and the web returned no results for this exact tool. It may be a private project, an early-stage effort that was not widely published, or the name/handle may be slightly different.

What the broader pattern looks like: There is a large ecosystem of tools implementing the "parallel agents with git worktrees" pattern. Notable open-source implementations include:

If more specific information about 999x_ai's project becomes available, this section should be updated.

Sources:

6. claude-flow -- Reuven Cohen (ruvnet)​

What it is: A multi-agent orchestration platform for Claude Code. Enables deploying coordinated agent swarms, autonomous workflows, and conversational AI systems. Connects via MCP (Model Context Protocol) to extend Claude Code with swarm capabilities. Can deploy 60+ specialized agents in coordinated swarms with self-learning, fault-tolerant consensus, and distributed intelligence.

Open source: Yes. MIT license. 14,300+ GitHub stars, 1,700+ forks.

Creator: Reuven Cohen (GitHub: ruvnet). Serial entrepreneur, Techstars mentor. Co-founder and CEO of AwardPool. Mentored startups that collectively raised $500M+. Also maintains agentic-flow (for deploying hosted agents with alternative low-cost models).

Business model: No paid tier. Entirely free and open source. No commercial component visible. Cohen appears to use claude-flow as a portfolio/reputation project that supports his broader consulting and entrepreneurial activities rather than as a direct revenue source.

Technical setup:

  • Installed via npm (npx claude-flow@latest) or curl
  • MCP server that extends Claude Code with orchestration tools
  • Swarm primitives: agent roles, topology config, message passing, consensus, execution backends
  • Task decomposition and role specialization
  • Lifecycle management and fault tolerance
  • RAG integration
  • SQLite-based persistent memory
  • Discord community for support
  • v1 released June 2025, v3 January 2026

Sources:

7. Crystal -- Stravu (now Nimbalyst)​

What it is: An Electron desktop application for running multiple Claude Code and Codex sessions simultaneously in parallel git worktrees. Launch multiple sessions with different prompts, compare results, selectively merge the best approaches back to main.

Open source: Yes. MIT license. 2,900+ GitHub stars.

Creator: Stravu (now apparently rebranded to Nimbalyst, as stravu.com/crystal redirects to nimbalyst.com/crystal). The specific founders/team members are not publicly disclosed on the website or GitHub. Crystal was built as an internal tool to speed up Stravu's own development, then open-sourced.

Business model: No paid tier. No commercial component mentioned. Entirely free and open source. Not affiliated with, endorsed by, or sponsored by Anthropic or OpenAI.

The likely play: Crystal appears to be a developer-tool marketing vehicle for whatever Stravu/Nimbalyst's primary product is. The rebrand to Nimbalyst and the redirect suggest the company is evolving beyond just Crystal. No pricing or paid features have been announced.

Technical setup:

  • Electron desktop app (cross-platform)
  • Git worktrees for session isolation
  • Each session gets its own working directory
  • Supports both Claude Code and Codex
  • Visual session management with comparison views
  • Merge workflows built in

Sources:

8. Compound Product / Ralph -- Ryan Carson (@ryancarson)​

What Ralph is: An autonomous AI agent loop that runs repeatedly until all PRD items are complete. Each iteration spawns a fresh AI instance with clean context. Memory persists through git history, progress.txt, and prd.json. The system picks incomplete stories from a JSON task list, implements them individually, runs quality checks, commits successful changes, and updates the status before repeating.

What Compound Product is: A self-improving product system built on top of the Ralph pattern. Reads daily performance reports (metrics, errors, feedback), identifies highest-impact issues, and autonomously implements fixes as PRs. Operates in 4 phases: Analysis (LLM reads reports, outputs prioritized JSON), Planning (creates PRD and granular tasks), Execution (loops through tasks with quality checks), Output (pushes to feature branch, creates PR).

Open source: Both are MIT license.

Supported agents: Amp CLI (default) and Claude Code (--tool claude).

Business model for Ralph/Compound Product: None. Both are free open-source tools with no commercial component.

How Ryan Carson actually makes money: He is the Builder in Residence at Sourcegraph for Amp (Sourcegraph's AI coding agent). Amp's founder Quinn recruited Carson after noticing his public engagement with the product. Carson also builds Untangle, an AI-powered divorce assistance platform (personal project inspired by family experiences). Previously founded:

  • DropSend (acquired 2008)
  • Carsonified (acquired 2011)
  • Treehouse (raised $23M VC, 100+ employees, 1M+ students, acquired 2021)

The Ralph pattern is named after "Ralph Wiggum" and was popularized by Geoffrey Huntley, then implemented/packaged by Carson. It has become a well-known pattern in the AI coding community. Carson publicly builds Untangle using Ralph+Amp as a real-world case study for Amp's capabilities, so Ralph serves as both a genuine tool and a marketing vehicle for Amp.

Technical details of Ralph:

  • Shell script (ralph.sh) as the main loop
  • PRD stored as JSON with user stories and acceptance criteria
  • Each iteration: fresh context window, picks next incomplete story, implements, runs tests/lint/type checks, commits if passing, updates progress
  • Memory: agents.md (long-term notes), progress.txt (iteration context), git history
  • Skills system for custom extensions

Sources:

9. 8090 Software Factory -- Chamath Palihapitiya​

What it is: An AI-native SDLC orchestration platform where PMs, designers, engineers, and QA collaborate to ship software. Positioned as the tool to "replace/rewrite all the legacy software in the world with modern alternatives." Started as an incubator (January 2024), evolved into a product platform (Software Factory launched ~September 2025).

Open source: No. Proprietary commercial product.

Founder: Chamath Palihapitiya. Funded entirely by him (not a traditional VC-backed startup). Social Capital (his fund) has $2B+ AUM. Chamath describes 8090 as "an experiment in high-efficiency software entrepreneurship."

Pricing:

PlanPriceDetails
Team$200/seat/monthFull platform access + standard support + token-based usage costs
EnterpriseCustom (volume-based)Priority SLAs, guided onboarding, volume discounts
Custom Factory LinesContact salesReverse engineering agents for legacy codebases, modernization support, includes 8090 Forward Deployment Engineers
Custom DeliveryStarting at $1M/yearFully managed SaaS -- 8090 handles hosting, maintenance, security, updates. Customer retains ownership of business logic

The promise: 80% of features at 90% less cost. Originally used AI + offshoring; now primarily AI-driven with the Software Factory product.

Technical setup -- four core modules:

  1. Refinery -- Collaborative requirement definition
  2. Foundry -- Architecture and system-level decision capture
  3. Planner -- Translates product intent into structured work orders
  4. Validator -- Converts feedback into actionable development tasks

Additional capabilities:

  • "Reverse engineering agents" that integrate with existing codebases to build knowledge graphs retroactively
  • "Assembly Lines" that memorize and automate specific patterns for repeatable execution
  • Can ingest entire legacy codebases for documentation/mapping/migration

Sources:

10. Boris Cherny (@bcherny)​

Who he is: Creator and Head of Claude Code at Anthropic. Built Claude Code as a side project in September 2024; it has since grown into one of Anthropic's flagship products. Previously worked at Meta (senior engineer on Chats in Groups). Was briefly recruited by Anysphere (Cursor) before being recruited back by Anthropic.

What he builds: Claude Code itself. He does not have a separate open-source tool or side business. His "product" is Claude Code, which is Anthropic's commercial AI coding agent.

How the 10-30 PRs/day translates to revenue: Boris's extreme productivity is a demonstration of Claude Code's capabilities, which directly drives Anthropic's revenue:

  • Claude Code achieved nearly $1 billion in annualized revenue within 6 months of launch
  • Anthropic's total revenue grew from $1B to $7B run-rate in 9 months
  • Revenue comes from enterprise paid subscriptions and commercial API usage
  • Boris landing 20-27 PRs/day with 100% AI-generated code is the ultimate product demo

His exact technical setup:

  • 10-15 concurrent Claude Code sessions simultaneously
    • 5 sessions in terminal (iTerm2 tabs, numbered 1-5, with OS notifications for when input needed)
    • 5-10 sessions in browser (claude.ai web interface)
    • Additional mobile sessions started in the morning, checked later
  • Uses a "teleport" command to hand off sessions between web and local terminal
  • Sticks to Opus 4.5 with thinking (slower but needs less hand-holding)
  • 10-20% session abandonment rate (sessions that go sideways)
  • Philosophy: "optimizing for throughput, not conversation" -- treats sessions like compute resources, schedules cognition, keeps sessions "hot"
  • Described as feeling "more like Starcraft" than traditional coding -- commanding autonomous units rather than typing syntax

Productivity stats (public):

  • Last 30 days: 259 PRs, 497 commits, 40K lines added, 38K removed
  • Every line written by Claude Code + Opus 4.5
  • Shipped 22 PRs one day, 27 the next
  • Roughly 100 PRs/week sustained

Business model: Boris is a salaried employee at Anthropic. His productivity is Anthropic's product marketing. "Pretty much 100%" of code across Anthropic is AI-generated according to him. He does not separately monetize his workflow or tools.

Sources:

Summary: Business Model Comparison​

ToolOpen SourceLicenseRevenue ModelPrice Range
dmuxYesMITNone (side project of FormKit creator)Free
SuperconductorNoProprietarySaaS platform fee (BYOK for LLM)Not disclosed; early access
TerragonYes (post-shutdown)Apache-2.0SaaS subscription (dead)Unknown; shut down Feb 2026
OpenClawYesMITCloud hosting ($39/mo planned) + enterprise supportFree self-hosted; $5-30/mo in API costs
swarm-codeUnknownUnknownUnknownUnknown
claude-flowYesMITNone (portfolio/reputation project)Free
CrystalYesMITNone (company marketing tool)Free
Ralph/CompoundYesMITNone (marketing for Amp/Sourcegraph)Free
8090 Software FactoryNoProprietarySaaS + managed delivery$200/seat/mo to $1M+/year
Boris ChernyN/AN/ASalaried at Anthropic; Claude Code is the productAnthropic API/subscription pricing

Key Patterns​

  1. Most agent orchestration tools are free/open source. dmux, claude-flow, Crystal, Ralph are all MIT-licensed with no commercial component. The creators make money elsewhere (FormKit, Sourcegraph, consulting, etc.).

  2. The real money is in the platform layer. Superconductor and 8090 are the only ones charging directly -- by providing cloud infrastructure, team collaboration, and enterprise-grade orchestration on top of the underlying AI agents.

  3. BYOK is the norm. Tools generally do not bundle LLM costs. Users bring their own Claude Max subscription or API keys. The tool adds orchestration/infrastructure value on top.

  4. Terragon's shutdown is a warning. Building a thin orchestration layer on top of Claude Code/Codex is vulnerable to the underlying providers adding the same features natively (Anthropic's background tasks, Claude Code Agent Teams, OpenAI Codex's built-in background mode).

  5. The biggest revenue comes from being the AI provider itself. Boris Cherny's 20+ PRs/day workflow is a product demo for Anthropic's $7B run-rate business. The orchestration tools are accessories; the model provider captures most of the value.

  6. Open-sourcing is a credibility/marketing strategy. Ralph promotes Amp. Crystal promotes Stravu/Nimbalyst. claude-flow promotes Reuven Cohen's consulting. The pattern is: give away the tool, monetize the attention or adjacent product.