Agent design is still hard
12 hours ago
- #AI Agents
- #Machine Learning
- #Software Development
- Building agents is messy with SDK abstractions breaking during real tool use.
- Explicit cache management is preferred for predictability and control, especially with Anthropic's approach.
- Reinforcement during the agent loop is crucial for guiding the agent and handling failures.
- Shared state via a file-system-like layer is essential for agent operations.
- Output tooling is tricky, with challenges in tone control and ensuring tool calls.
- Model choice depends on the task, with Haiku and Sonnet being preferred for tool calling.
- Testing and evaluations are challenging due to the agentic nature of the systems.
- Amp is being trialed for its innovative agent design and sub-agent interactions.