Someone is finally bringing order to AI agent tool calling.

On August 23, Amazon Web Services announced it is open-sourcing a project called Dogwood, positioned as "tool calling specifications and a reference implementation for AI agents." Put simply: it creates a unified standard and open-source implementation for how agents call tools, pass parameters, handle errors, and manage permissions.

The context here is that AI agent tool calling is currently a mess. Every framework and every model vendor has its own tool calling format — OpenAI has function calling, Anthropic has tool use, MCP is one protocol, A2A is another. Developers building agents that run across multiple platforms have to adapt to half a dozen different formats. It's painful.

What Problems Does Dogwood Solve?

Dogwood has three core goals: unified format, standardized error handling, and normalized permission management.

First, unified tool description format. Regardless of which model or framework you're using underneath, tool descriptions, parameter schemas, and return value formats are all defined using the same specification. Dogwood defines a JSON Schema-based tool description language that supports type checking, parameter validation, and automatic documentation generation. That means developers write tool definitions once, and they work across all Dogwood-compatible agent frameworks.

Second, standardized error handling. Right now, when an agent's tool call fails, every framework returns errors differently, has different retry logic, and feeds errors back to the model differently. Dogwood defines standard error codes, error categories, and retry policies — which errors are retryable, which require rethinking, which should terminate the task — all with clear guidance.

Third, normalized permission management. This is the most dangerous gap in agents today: an agent with your API key might call dangerous tools (like deleting a database, transferring money, sending emails) without your knowledge. Dogwood includes a built-in permission declaration and approval mechanism. Tools declare what permissions they need; sensitive tool calls require human confirmation; all operations have audit logs.

Another Player in the Protocol Wars

Dogwood isn't the first player in this space. Before it, there were two major protocols competing: MCP (Model Context Protocol) for connecting agents to tools and data, and A2A (Agent-to-Agent Protocol) for inter-agent communication. Dogwood focuses more narrowly on tool calling standardization itself, and can be seen as complementary to MCP on tool definition and error handling.

AWS choosing to open-source Dogwood now sends a clear signal: the agent protocol layer is becoming a new battleground for big tech. Whoever sets the standard controls the gateway to the agent ecosystem. Just like the OpenStack vs. Cloud Foundry competition in the cloud era, or Android vs. iOS in the mobile internet era — every paradigm shift brings a standards war.

The question now is: will all these protocols eventually unify, or will fragmentation persist? Historical precedent suggests "fragment first, converge later" — early on everyone does their own thing, then market competition gradually narrows it down to 1-2 mainstream standards, possibly unified under an open-source foundation. A2A has already been donated to the Linux Foundation; MCP is on a similar path; Dogwood, as an AWS open-source project, could follow.

For developers, the short term may mean "multiple protocols coexisting" chaos, requiring adaptation across standards. But in the long run, having standards is better than no standards. Big companies investing resources in standardization is good for the industry as a whole.

For agents to truly reach production environments, standardization is a hurdle that must be crossed. Dogwood is Amazon throwing a stone into the pond — and the waters will keep stirring.

See you tomorrow.