MULTI-AGENT WORKFLOW ARCHITECTURE:

PATTERN CHOICE:

1. SEQUENTIAL (Agent A → B → C, output of A feeds B)
   → Best for: Reasoning chains (research → analyze → write)
   → Context: Pass full previous output + 1-line task summary
   → Failure: If B fails, retry B; don't restart A
   → Example: Research agent gathers facts → Analyst scores them → Writer creates article

2. PARALLEL (Agent A, B, C all run, results merged)
   → Best for: Independent analyses (code review + security scan + performance test)
   → Context: Each agent gets same input, merge-step synthesizes
   → Failure: If 1 agent fails, continue with 2 results (mark as partial)
   → Cost: 3x tokens (3 agents running), but 3x faster wall-clock time

3. HIERARCHICAL (Manager → delegates specialized agents)
   → Best for: Complex problems (manager decides which expert to call)
   → Context: Manager sees agent outputs, decides next step
   → Failure: Manager failure = whole pipeline blocked (single point)
   → Example: Manager reads code → calls security expert OR performance expert based on risk

HANDOFF CONTRACTS:

→ Between agents: Define expected input shape (JSON schema)
→ Track context tokens per agent (prevent explosion)
→ Agent 1 output: Explicit pass/fail + data or error reason
→ Error case: If Agent 2 can't parse input, return structured error, don't hallucinate

FAILURE HANDLING:
• Retry logic: 2 automatic retries on transient errors (network, timeout)
• Fallback: If Agent A fails 3x, skip and pass raw data to Agent B (marked as incomplete)
• Logging: Every handoff logged (timestamp, tokens used, success/failure)

CONTEXT BUDGET:
• Total: 80K tokens for full pipeline
• Per agent: 20K max (prevents runaway)
• Compression: Summarize mid-results before next agent (cut irrelevant detail)

COST MONITORING:
→ Track cost per workflow run (e.g., code review = 0.85 cents average)
→ Alert if single run exceeds 2x average (signal of infinite loop or bug)