Claude provider: agentic loop `max_iterations=10` too low and not configurable

[franklixuefei]

Claude provider: agentic loop max_iterations=10 too low and not configurable

Problem

The Claude provider's _execute_agentic_loop() has a hardcoded max_iterations=10 default. This is the number of tool-use roundtrips a single agent can make before the provider throws a ProviderError. For any agent that needs to read multiple files, write code, and verify results, 10 iterations is far too few.

The Copilot provider has no equivalent iteration limit — it uses the SDK's built-in agentic loop with a 30-minute wall-clock timeout (max_session_seconds=1800) and a 5-minute idle timeout (idle_timeout_seconds=300). There is no cap on the number of tool calls.

Impact

Complex agents (e.g., a "coder" agent that reads a plan, explores the codebase, writes implementation code, and verifies changes) routinely need 20-40+ tool-use iterations. With the current limit of 10, these agents always fail with:

Agentic loop exceeded maximum iterations (10)
💡 Suggestion: The agent may be stuck in a tool-use loop. Check your MCP tools.

This makes the Claude provider unusable for any non-trivial coding workflow.

Root Cause

In conductor/providers/claude.py, _execute_agentic_loop():

async def _execute_agentic_loop(
    self,
    messages: list[dict[str, Any]],
    model: str,
    temperature: float | None,
    max_tokens: int,
    tools: list[dict[str, Any]] | None,
    output_schema: dict[str, OutputField] | None,
    has_output_schema: bool,
    max_iterations: int = 10,  # <-- hardcoded, not configurable
    ...
) -> tuple[ClaudeResponse, int | None, bool]:

The value is never overridden — the caller in _execute_with_retry() doesn't pass it, and there's no way to set it from the workflow YAML runtime config or per-agent config.

Suggested Fix

Option A: Make it configurable (preferred)

1. Add a max_agent_iterations field to the Claude provider's runtime config (similar to max_tokens, temperature, etc.)
2. Allow per-agent override via the agent config in the workflow YAML
3. Set a reasonable default (e.g., 50) that matches the Copilot provider's practical behavior

Example workflow YAML:

runtime:
  provider: claude
  config:
    max_agent_iterations: 50  # or per-agent override

Option B: Match Copilot provider behavior

Replace the iteration count limit with a wall-clock timeout, matching the Copilot provider's IdleRecoveryConfig pattern:

• max_session_seconds: 1800 (30 minutes per agent)
• idle_timeout_seconds: 300 (5 minutes without API activity)

This would provide true feature parity.

Option C: Quick fix (minimum viable)

Just increase the default from 10 to 50. This covers most practical use cases without architectural changes.

Workaround

Patch the default locally:

sed -i 's/max_iterations: int = 10,/max_iterations: int = 50,/' \
  "$(python -c 'import conductor.providers.claude; print(conductor.providers.claude.__file__)')"

Environment

• conductor-cli: installed from git+https://github.com/microsoft/conductor.git
• Provider: claude
• Workflow: implement.yaml (coder agent with filesystem MCP server)
• The coder agent typically makes 20-40 tool calls per epic implementation