Feature: Per-Agent Retry Policies

[jrob5756]

Summary

Add configurable retry policies at the agent level so transient failures (API errors, rate limits, timeouts) don't kill entire workflows.

Motivation

Production agentic AI demands resilience. Today, one transient API failure kills the entire conductor workflow — the only recovery is checkpoint/resume (all-or-nothing). Production frameworks like LangGraph are positioned as leaders partly due to configurable retry, and real-world incidents (Amazon retail outages from AI agents, "two AIs talked for 2 hours booking nothing") underscore the need for fault tolerance.

Proposed Design

agents:
  - name: analyzer
    model: gpt-5.2
    retry:
      max_attempts: 3              # default: 1 (no retry)
      backoff: exponential         # or "fixed"
      delay_seconds: 2             # base delay
      retry_on:
        - provider_error           # API 500s, rate limits
        - timeout                  # agent-level timeout exceeded
        # NOT: validation_error — output schema mismatches indicate logic bugs, not transience

Behavior

• Retry counter resets per agent execution (not per workflow)
• Exponential backoff: delay * 2^attempt with jitter
• Events emitted on each retry attempt (agent_retry event)
• Final failure after exhausting retries follows existing error handling (fail_fast, continue_on_error, etc.)
• Works in both sequential and parallel execution contexts

What Does NOT Retry

• Output validation errors (schema mismatches) — these indicate prompt/schema issues, not transience
• Human gate timeouts — these are intentional pauses
• Workflow-level timeout exceeded — hard stop

Why It Fits Conductor

• Declarative YAML config, no imperative code needed
• Per-agent granularity, not per-workflow — fine-grained control
• Pairs naturally with existing fail_fast/continue_on_error parallel semantics
• Pairs with agent-level timeouts (Feature: Agent-Level Timeouts #82) and fallback models (Feature: Cost Budgets & Adaptive Model Routing #84) for a complete resilience story

Effort Estimate

Low — wraps existing AgentExecutor.execute() in a retry loop with backoff logic.