Comparison Matrix

Pull the three provider envelopes side by side and the matrix gets short fast. Most differences are cosmetic; a few are real.

Concept	OpenAI	Anthropic	Gemini
Tool list field	`tools`	`tools`	`tools` (in config)
Schema field	`parameters`	`input_schema`	`parameters` (Schema object)
Args type back	JSON string	dict	dict
Tool choice	`"auto"\|"required"\|spec`	`type: "auto"\|"any"\|"tool"`	FunctionCallingMode enum
Stop signal	`response.completed` + `output[].type`	`stop_reason`	presence/absence of `function_call` part
Streaming tool args	Deltas	Deltas (per content block)	All-at-once
Pre-built native tools	web_search, code_interpreter, file_search	computer_use, text_editor, bash	code_execution, google_search
MCP integration	via Apps SDK	via mcp_servers in Messages API	via Vertex AI tool config

The four cells where the difference is real, not cosmetic, are: the args type (parse vs use), the streaming behavior (deltas vs all-at-once), the pre-built tools (each provider has unique offerings), and the MCP integration shape. Build a thin translation layer for the cosmetic ones and accept that the real ones require provider-specific code.

The temptation is to write a giant abstraction that pretends all three are the same. Resist it for as long as you can. The shape of "send messages and tools, get tool calls back, send results in" is already a small abstraction; over-abstracting hides the real differences and produces SDK-like layers nobody else wants to maintain.

Code

Tiny portability shim — only what you actually need·python

def normalize_tool_calls(provider, response):
    if provider == "openai":
        return [(item.name, json.loads(item.arguments)) for item in response.output if item.type == "function_call"]
    if provider == "anthropic":
        return [(b.name, b.input) for b in response.content if b.type == "tool_use"]
    if provider == "gemini":
        return [(p.function_call.name, dict(p.function_call.args))
                for p in response.candidates[0].content.parts
                if p.function_call]
    raise ValueError(f"Unknown provider {provider!r}")

Exercise

Take the matrix above and shrink it to the four 'real differences' rows. Pin those four rows above your editor (or in a comment at the top of your loop module). Whenever you find yourself writing a portability hack, check if it is one of those four. If yes, accept the divergence. If no, your shim is leaking too much.

Code

External links

Exercise

Progress

Comments 0