Why I'm Betting Against AI Agents in 2025 (Despite Building Them)
9 months ago
- #Production Challenges
- #AI Agents
- #Tool Engineering
- 2025 is widely hyped as the year of AI agents, but practical experience shows significant challenges.
- Error rates compound exponentially in multi-step workflows, making high reliability difficult to achieve.
- Context windows create quadratic token costs, making long conversations economically unsustainable.
- The real challenge is designing tools and feedback systems that agents can use effectively.
- Production-grade agent systems require bounded contexts, verifiable operations, and human decision points.
- Successful agents are often stateless, focused tools rather than conversational systems.
- Tool engineering is critical, requiring careful design to provide effective feedback without overwhelming context.
- Integration with real-world systems is messy and often where AI agents fail.
- Successful agent systems combine AI for complexity with human control and traditional engineering for reliability.
- The future of AI agents lies in constrained, domain-specific tools rather than fully autonomous systems.