World's Fair 2026
Agents Need Feature Flags - Sachin Gupta
Overview
This talk argues that AI agents, capable of performing high-impact actions like sending money or modifying databases, are being deployed without the safety infrastructure common in traditional web development. It introduces six specific types of feature flags—prompt variants, tool access, model routing, memory policy, autonomy level, and kill switches—that are necessary to manage the unique behavior surfaces of agents and prevent incidents. The core thesis is that applying established web development disciplines, particularly robust kill switches and phased rollouts, is crucial for controlling and safely iterating on AI agent functionality.
Who should watch
- AI Engineers building and deploying agents.
- Product Managers overseeing AI agent features.
- Builders seeking to implement safer agent development practices.
- Teams experiencing unexpected agent behavior or incidents.
- Anyone concerned with the operational safety of AI systems.
Key takeaways
- AI agents require specialized feature flags beyond simple on/off toggles, addressing six distinct behavior surfaces: prompts, tools, models, memory, autonomy, and sub-agents.
- A kill switch, capable of taking effect in seconds without deployments, is the most critical safety mechanism and should be implemented first.
- Phased rollouts, similar to canary releases in web development, are essential for testing new prompts, tools, or model changes on a small percentage of users before full deployment.
- Incidents like database deletion by an agent or costly infinite loops highlight the risks of deploying powerful agents without adequate control mechanisms.
- Tool access flags are vital for managing agents that interact with sensitive operations like financial transactions or data modification.
- Model routing flags allow for quick migration or fallback to different models during outages or deprecations, preventing complete agent failure.
- Memory policy flags are crucial for managing data privacy, agent consistency, and compliance with regulations like GDPR.
- A rollout playbook emphasizes shipping kill switches first, wrapping tools with flags, staging autonomy levels, moving prompts to configurable variants, and diligently monitoring key metrics.
Notable quotes
*We are shipping them the way web team used to ship in 2008.*
*The boring infrastructure that keep deploy safe is already a solved problem.*
*Ship the kill switch first. If you do nothing else, give your agent one agentwide kill switch and one per tool kill switch.*
Unofficial community note. Prefer the recording for nuance.