News · 2026-07-01

'agency-agents' packages 150+ role-playing AI agents into one open-source 'AI agency'

The open-source project agency-agents packages more than 150 specialized AI agent personas into a single library designed to function like a complete digital agency. Organized into over 13 professional divisions -- engineering, design, finance, marketing, game development, and more -- each agent comes with a defined personality, a working process, and expected deliverables, and the whole collection is built to run multi-agent workflows natively inside Claude Code and other agentic coding tools. It has drawn a large and fast-growing following on GitHub.

Key facts

• agency-agents defines 150-plus specialized AI agents across more than 13 professional divisions, each with a personality, process, and required output.
• It is built to work natively with Claude Code and includes conversion and install scripts for other agentic coding tools.
• The repository has attracted a very large star count and shows active, frequent commits.
• Primary sources: the GitHub repository (msitarzewski/agency-agents) and the project site at agencyagents.dev.

The idea builds on a pattern that has taken hold across the agent ecosystem: instead of one general assistant trying to do everything, you assemble a team of narrow specialists, each with a focused role, and let them collaborate. agency-agents takes that to its logical conclusion by shipping an entire org chart. There are 'Frontend Wizards' for interface work, 'Reddit Community Ninjas' for community marketing, 'Whimsy Injectors' for polish, and -- tellingly -- 'Reality Checkers' whose whole job is to push back on the other agents. The personas are playful, but the underlying structure is serious: give each agent a bounded remit and a defined deliverable, and complex projects decompose into coordinated pieces.

Mechanically, each 'agent' is not a standalone program but a definition -- a bundle of prompt, context, and expected behavior -- that runs on top of an existing agentic coding tool. That is why native Claude Code support matters: the framework leans on the host tool for actual execution and supplies the specialization on top. The included conversion scripts let the same definitions run on other major agent platforms, which is a bet on interoperability in a market where no single agent tool has won.

An analogy: it is less a piece of software and more a staffing agency's playbook -- a filing cabinet of pre-written job descriptions, processes, and quality bars that you drop into whatever office (agent tool) you already run.

Why it matters: the project is a snapshot of where the AI agents field's energy is right now -- role specialization, multi-agent collaboration, and reusable agent definitions shared as open source. The inclusion of dedicated 'Reality Checker' and reviewer agents is a quiet acknowledgment of a hard-won lesson: multi-agent systems need built-in skepticism, or they confidently amplify each other's mistakes. That the framework is portable across tools, rather than locked to one, reflects a maturing ecosystem where prompt-and-process assets are becoming a distinct layer above any single vendor.

The honest caveat: a library of 150-plus personas is impressive to browse, but breadth is not the same as reliability, and the hard part of multi-agent systems has never been defining roles -- it is coordination, error propagation, and cost. More agents means more calls, more tokens, and more chances for one agent's mistake to cascade, which connects directly to the week's cost stories like Meta capping employee AI token spend. Reported star counts and commit activity signal real interest, but they measure attention, not whether these workflows deliver production results better than a single well-prompted agent. The concept is genuinely useful as a starting kit; treat the '150 agents' framing as a menu to draw from, not a promise that running all of them at once is a good idea. Follow the agent ecosystem at Ground Truth.

Primary source, verified: read the paper →

Topics: ai-agents · open-source · multi-agent · claude-code · workflows