Team Topologies for AI Systems: Building Organizational Structures for Agentic AI
This article discusses how Team Topologies principles can provide the 'infrastructure for agency' needed for successful AI investments, addressing organizational rather than purely technical hurdles. It emphasizes using bounded agency and stewardship to govern AI agents, much like human teams, and introduces an 'Innovation and Practices Enabling Team' for knowledge diffusion.
Read original on InfoQ ArchitectureThe article highlights that many businesses struggle to gain tangible returns from AI investments due to organizational rather than technical obstacles. Matthew Skelton, co-author of Team Topologies, posits that the framework's principles offer the necessary structure for managing AI agents effectively, especially in light of vulnerabilities like Excessive Agency (LLM06) in the OWASP Top 10 for LLM Applications.
Bounded Agency and Stewardship
A core concept is bounded agency, where the authority of both human and AI agents is intentionally constrained by rules and guardrails. This ensures that delegated initiatives remain governable, clarifying mission and focus within the organization. The article argues that organizations already structured for bounded agency in human teams will find the transition to agentic AI systems significantly smoother. This contrasts with the common pitfall of granting AI tools unbounded access to data, creating security vulnerabilities and compromising domain fidelity.
Excessive Agency Risk
Unbounded access for AI agents to data resources is a critical security vulnerability (OWASP LLM06). Companies should question why they would grant an AI agent write access that would be denied to a human, emphasizing the need for strict security boundaries and protecting domain fidelity to maintain business intent.
The concept of stewardship is presented as a more productive framing than mere ownership. It encourages teams to look after systems for future contributors, fostering a long-term perspective. This aligns with managing AI context windows, ensuring agents operate within defined boundaries to maintain coherence and prevent 'hallucinations,' similar to how human cognitive load is managed.
Innovation and Practices Enabling Team
To scale successful patterns, Skelton introduces the Innovation and Practices Enabling Team. This specialized team identifies successful organizational patterns and facilitates their diffusion across other departments. This approach fosters active knowledge diffusion, building momentum through shared success rather than mandatory compliance, which is crucial given technology's rapid evolution. Examples like JP Morgan's reduction of dependencies in its Athena platform via an opt-in model ('friendly FOMO') illustrate this strategy.
- Bounded Agency: Intentional constraints on authority for governability.
- Stewardship: Long-term care for systems over simple ownership.
- Innovation and Practices Enabling Team: Facilitates knowledge sharing and best practice adoption.