The open-source project “The Agency” has released over 230 pre-built AI agents, attracting explosive attention among engineers and business leaders. The background is the rapid shift in demand from general-purpose chat AI to “expert AI agents” specialized in specific business processes and deliverables.

Development innovations brought by a team of over 230 experts

The repository “agency-agents” (also known as The Agency), published by Msitarzewski, is a collection of over 230 AI agents specialized for specific roles. Born from a Reddit thread, this project grew after months of iterative development to build a vast team of experts across 16 departments including engineering, design, marketing, and sales.

Each agent is not just a prompt template but is characterized by its unique identity, expertise, specific workflows, and success metrics to achieve. For example, it covers technical roles like front-end development and back-end architect, Reddit community builders, and even unique roles like ‘Whimsy Injector.’

As of July 2026, this repository has over 135,000 stars on GitHub, indicating strong interest in “pre-built agents” that can instantly handle specific tasks. Users can now “assign” the necessary experts to their projects, as if forming their ideal team. Please refer to the table below.

Seamless integration environment with leading AI tools

Another strength of Agency Agents is their high affinity for the main AI coding tools and platforms developers use daily. This project provides native desktop apps for macOS, Linux, and Windows, allowing users to install agents in tools like Claude Code, Cursor, GitHub Copilot, and Osaurus with a single click.

For engineers who prefer terminal operations, there are interactive installers with automatic detection features and scripts that can handle multiple tools in parallel. Additionally, in Claude Code, simply copying the agent file from the repository to a specific configuration directory allows you to call the agent directly in conversation and delegate tasks.

Furthermore, as of mid-2026, a trend is the emergence of derivative projects like “agency-orchestrators,” where these specialized agents are orchestrated (coordinated) and multiple AI experts can coordinate from a single instruction to create complex plans within minutes. In this way, an ecosystem is rapidly taking shape that combines agents with individual skill sets to automate larger workflows.

Deliverable-Focused Design and the Power of Open Source

Design focused on “deliverables” that break down business operations

What sets Agency Agents apart from existing prompt collections is their design philosophy. Each agent is designed with a focus on ‘what to deliver’ rather than ‘what they will become.’ The agent definition file describes not only identity and personality but also specific missions, domain-specific key rules, and technical deliverables including code examples.

For example, the “Evidence Collector” agent does not simply test the code but has a strict process that identifies 3 to 5 issues by default and requires visual proof such as screenshots for every point. In this way, the definition of a “Code of Conduct” that enables AI to perform business processes with equal or greater accuracy than humans enhances its practicality in commercial environments.

Use cases are also envisioned in detail, such as scenarios where a startup’s MVP is built by combining front-end developers, back-end architects, growth hackers, and reality checkers to bring the product to market in the shortest possible time. Value creation through this “combination of expertise” is a domain that conventional general-purpose AI has struggled to reach.

The power of open source that anyone can customize

This project is published under the MIT license, allowing free use, modification, and redistribution for both personal and commercial purposes. Users can easily fork existing agents and adjust them to fit their internal policies or specific project rules. This “transparency” and “customizability” are precisely what make Agency Agents a major appeal that black-boxed commercial AI tools lack.

Additionally, multilingual support and localization by the community are actively being carried out. In the Japanese environment, volunteers maintain a Japanese repository that includes 281 agents localized for Japan and 97 agents original to the Japanese market. This enables immediate access to specialized agents tailored to Japanese business customs and language-specific nuances.

Furthermore, a contribution cycle has been established where users share successful agent configurations on GitHub Discussions or add new agents via PR (pull requests), continuously evolving the entire project every day. This open development model ensures that the latest “expertise” is constantly replenished in line with advances in AI technology.

The Importance of Security and Compliance

Security “pitfalls” lurking in autonomous agents

As AI agents gain advanced privileges and can autonomously access file systems and networks, new security risks are also emerging. In February 2026, the Dutch Data Protection Authority (AP) issued a warning that autonomous AI agents could become “Trojan horses.” In particular, ‘indirect prompt injection,’ which executes hidden malicious instructions when processing information from external sources (websites or emails), poses a serious threat.

IBM Research’s 2025 penetration test results revealed that defenses trusted by developers, such as “user approval” and “allowlists by string matching,” can be easily bypassed by advanced injection techniques such as markdown hiding. The risk of “supply chain attacks,” where code generated by AI agents introduces potential vulnerabilities to downstream projects, cannot be ignored.

When using powerful collections of agents like Agency Agents, it is essential to run these in a sandboxed environment and build a “defense-in-depth” system that restricts network access. Granting permissions to tools should be based on the “principle of least privilege” and limited to what is truly necessary for the work.

Compliance with the EU AI Act approaching August 2026

2026 marks a milestone year as AI regulation shifts from “guidelines” to “legally enforceable enforcement.” In particular, the requirements for “high-risk AI systems” under the EU AI Act will be gradually implemented starting August 2026. AI agents that autonomously perform personnel evaluations, credit scoring, financial decisions, or critical infrastructure operations will be subject to strict management under this regulation.

When companies incorporate tools like Agency Agents into their operations, they should not simply introduce them as convenient tools, but must meet the following four legal requirements.

• Preparation of detailed technical documentation explaining decision-making logic

• Ensuring “open-loop” operations that allow external monitoring

• Establishment of human intervention points (human in the loop)

• Implementation of stop/overwrite mechanisms in case of unpredictable behavior

Since implementing these governances retroactively is costly and time-consuming, designing the architecture with a “governance-first” spirit from the initial stages is the only way to avoid future legal risks.

Future Outlook and Highlights

The emergence of Agency Agents symbolizes AI’s evolution from “conversation partners” to “teams that get the job done.” Going forward, automation of “multi-agent collaboration,” where multiple specialized agents autonomously interact and review deliverables, will accelerate even further, not only improving the accuracy of a single agent.

Furthermore, as market data for Q1 2026 shows, the criteria for evaluating AI agents have decisively shifted from “intelligence” to “software reliability and reproducibility.” The demand for a collection of compliant agents specialized in specific industries (such as healthcare, legal, and finance) will continue to grow. We cannot take our eyes off how much the “democratization of expertise” pioneered by Agency Agents will accelerate corporate DX.

Reference Page

• 【msitarzewski/agency-agents – GitHub】https://github.com/msitarzewski/agency-agents

• 【Lessons from Penetration Tests on Large-Scale Agent Systems – arXiv】https://arxiv.org/abs/2501.05405

• 【EU AI Act Text – Official】https://eur-lex.europa.eu/eli/reg/2024/1689/oj

[#AIエージェント #AgencyAgents #GitHub #生成AI #自動化 #オープンソース #エンジニアリング #AIガバナンス]