Batching tool calls with CodeMode¶

When the agent has many tools — especially a large drf-mcp bridge — each call is its own model round-trip, and every tool's schema bloats the prompt. pydantic-ai-harness's CodeMode capability collapses them into a single sandboxed run_code tool: the model writes Python that calls the tools as functions (with loops, asyncio.gather, intermediate variables) and gets the result in one round-trip.

This is a composition of two things django-ag-ui already gives you — the in-process drf-mcp bridge and the CAPABILITIES seam — plus the optional [harness] extra.

Install¶

pip install "django-ag-ui[harness]"

The [harness] extra pulls pydantic-ai-harness and its sandbox (pydantic-monty). The core install stays django + pydantic-ai-slim — the harness is lazy, only imported by the capability you wire in.

Wire it in¶

CAPABILITIES takes dotted paths to zero-argument callables that return a capability. Add one that returns a CodeMode:

# myproject/agent.py
from pydantic_ai_harness import CodeMode


def code_mode():
    return CodeMode()  # wraps every tool the agent has into `run_code`

# settings.py
DJANGO_AG_UI = {"MODEL": "anthropic:claude-sonnet-4.6"}

# urls.py
from myproject.agent import code_mode
from myproject.mcp import server as mcp_server

agent = AGUIServer(registry, drf_mcp_server=mcp_server, capabilities=[code_mode])

Now the agent exposes one run_code tool instead of the individual DRF tools; the model batches its DRF work into a single sandboxed script.

Wrapping only some tools¶

Pass a name list or a predicate to tools= to split the toolset — matching tools go behind run_code, the rest stay as direct tool calls. This is useful to keep, say, a destructive tool a ToolGuard gates as a direct call (so its approval interrupt still fires) while batching the read-only ones:

def code_mode():
    # Batch everything except the approval-gated delete.
    return CodeMode(tools=lambda ctx, td: td.name != "projects.delete")

Typed stubs — keep drf-mcp's output schema on¶

CodeMode renders each wrapped tool as a typed Python stub so the model knows what each call returns. A tool with no return schema renders -> Any and CodeMode emits a warning — the model loses that type information.

django-ag-ui's drf-mcp bridge carries drf-mcp's outputSchema onto the tool's return_schema, so bridged tools render typed as long as drf-mcp advertises the output schema — which it does by default (REST_FRAMEWORK_MCP["INCLUDE_OUTPUT_SCHEMA"] = True). Keep it on (and give your services an output serializer) for the best CodeMode experience; a service with no output serializer advertises no schema and its stub falls back to -> Any.

Note

CodeMode is one of several pydantic-ai-harness capabilities that drop into the same CAPABILITIES seam (compaction, step-persistence, subagents, …). They all ride the [harness] extra.