Agent Autonomy Tiers

What it is

Agent Autonomy Tiers formalizes the spectrum of control given to AI agents, ensuring that their operational leash matches the potential blast radius of their actions. This isn't about whether an agent can perform a task, but rather under what conditions it's permitted to act independently. We categorize agent actions into distinct tiers, from fully internal, unreviewed operations to external, high-impact actions that demand explicit human approval. This structure is critical for maintaining control and safety as agents become more capable and integrated into our product workflows.

Why it matters

As we build more sophisticated agents, particularly those leveraging advanced models like Claude Code or Gemini for complex reasoning and Tool Use Pattern, the risk profile of their actions increases. Unchecked autonomy can lead to unintended consequences, from minor data inconsistencies in Firebase to public-facing errors in a deployed product. By defining clear autonomy tiers, we minimize these risks while still capturing the efficiency gains of automation. It allows us to experiment with agent capabilities internally without fear of external impact, and to gradually increase autonomy as trust and validation build. This framework is a core component of our Human-in-the-Loop strategy, ensuring that human oversight is applied where it matters most, without becoming a bottleneck for low-risk operations.

How TV applies it

At Total Ventures, our agents operate across several defined tiers. For instance, an agent tasked with refactoring internal utility functions within a project's codebase (e.g., optimizing a data fetching routine or cleaning up unused imports) typically operates at a high autonomy tier. Its actions are confined to our internal development environment, and any issues are caught during standard CI/CD processes before deployment to Vercel. These agents often work with Structured Output via Zod to ensure their code changes adhere to predefined schemas.

Conversely, an agent generating content for Inky, our AI writing assistant, operates at a lower autonomy tier for public-facing outputs. It drafts content based on user prompts, but the final output is always presented to the user for review and approval before it's considered final. For tasks like sending transactional emails via Resend or making changes to live customer data, the agent generates a proposed action, which then requires explicit human review and approval through a dedicated internal interface. This tiered approach allows our internal agents to iterate rapidly on code and data structures, while ensuring that customer-facing interactions and product deployments remain under direct human supervision.

Common failure modes

A common failure mode is misassigning an agent to an inappropriate autonomy tier. Granting too much autonomy to an agent performing external actions can lead to public errors, customer dissatisfaction, or even data integrity issues. Forgetting to implement a robust Human-in-the-Loop step for high-impact actions is a critical oversight. Another failure mode is the 'slippery slope' where an agent's autonomy is gradually increased without re-evaluating its potential blast radius. What starts as a simple internal task can evolve into something with external implications if its capabilities expand without a corresponding adjustment to its autonomy tier. Finally, a lack of clear logging and audit trails for agent actions, regardless of tier, makes it difficult to diagnose issues or understand an agent's decision-making process when something goes wrong.