The Rabbitory
Four major labs shipped an agent framework or SDK update this month, all chasing the same goal: make it easy to wire a model up to tools and let it run multi-step tasks on its own.
The frameworks differ in details — how they handle state, how they sandbox tool calls, how they recover from errors — but they're converging on the same rough shape.
What's converging
Structured tool definitions, a planning/execution loop, and some form of persistent memory across steps are now table stakes rather than differentiators.
The framework war is really a developer-experience war — the underlying capability is close to identical across providers.
Where they still differ
- Error recovery and retry behavior when a tool call fails
- How much of the planning loop is exposed vs. hidden from the developer
- Pricing for the extra reasoning tokens agent loops burn through
For builders, the practical advice hasn't changed much: pick the framework that fits your existing stack, and budget for more tokens than a single-turn chat would use.