CrewAI and AutoGen

Agents

Each agent has a role (title and purpose), a goal, and a backstory. Agents can be assigned tools, a specific LLM, and reasoning depth configuration. The role/goal/backstory structure shapes how the LLM plays the specialist.

Tasks

Discrete work units with a description, expected output, and agent assignment. Tasks can receive context from previous tasks. Expected output definition is the most important task configuration - clear definitions produce more reliable results.

Sequential Process

Tasks execute in defined order. Each task output feeds the next. Predictable, testable, reliable. The recommended choice for production. 100+ built-in tools available out-of-the-box.

Hierarchical Process

A manager agent dynamically assigns tasks to workers. More flexible - but the manager frequently executes tasks itself rather than delegating, negating the architecture. Known production reliability issues require significant workarounds.

Hierarchical Delegation Failures

Manager agents default to executing tasks themselves rather than delegating. Known issue tracked in the GitHub repository. Requires custom manager prompting workarounds that make “hierarchical” less plug-and-play than advertised.

Our approach: Use sequential process in production. Avoid hierarchical until delegation reliability improves or you have the guardrail budget to compensate.

Cost Spirals from Undetected Loops

Without built-in cost monitoring, crews enter infinite retry loops that run for days. One documented case: $127/week → $47,000/month over four weeks due to an 11-day undetected loop.

Our approach: Implement hard iteration limits, cost alerting, and circuit breakers before any production deployment. Non-negotiable.

Hallucination Poisoning in Integrated Systems

When CrewAI agents integrate with external systems (JIRA, Salesforce, GitHub), hallucinated outputs from one agent are committed to memory and passed downstream as facts - creating cascading failures.

Our approach: Validate tool outputs at every handoff. Use structured Pydantic schemas to constrain what agents can assert.

Task Execution Freezes

Tasks get stuck in “THINKING” state indefinitely. No identified pattern, no recovery mechanism. Causes complete crew failure. Reported as GitHub issue #2997 (June 2025), unresolved.

Our approach: Implement timeout handling at the orchestration layer. Design for recovery rather than assuming tasks always complete.

AssistantAgent

LLM-powered reasoning agent. Generates responses and code suggestions but does not execute code by default. The reasoning engine in most AutoGen workflows.

UserProxyAgent

Represents the human interface. Can execute code blocks generated by AssistantAgent. The canonical AutoGen pattern: AssistantAgent generates code, UserProxyAgent executes it and returns results.

GroupChat

Multiple specialized agents in a shared conversation, each contributing expertise across multiple rounds. GroupChatManager orchestrates turn-taking until the group reaches a result.

Maintenance Mode

AutoGen is merging with Semantic Kernel into Microsoft Agent Framework. Significant new features are not being added. Consider the longevity implications before building production systems on AutoGen.

Infinite Conversation Loops

Agents get stuck in loops that don't terminate - including in basic quickstart configurations. Persistent across multiple major versions. The “gratitude loop” with GPT-3.5-turbo: agents express thanks indefinitely after completing a task because no termination condition triggers.

Our approach: Implement explicit termination conditions for every agent. Maximum iteration limits are non-negotiable. Test termination behavior across diverse inputs.

Token Cost Explosion

A 10-cycle Reflexion loop consumes ~50x the tokens of a single linear pass. Three-agent workflows average 8,000 tokens at GPT-4 pricing. Without cost controls, AutoGen workflows become economically unviable quickly.

Our approach: Set per-workflow cost budgets. Use OpenAIWrapper.print_usage_summary() to track costs per agent. Alert on cost anomalies at 5x expected baseline.

Context Window Overflow

AutoGen conversations accumulate history until they exceed context window limits. InvalidRequestError surfaces as a production failure. Context management tools exist (ListMemory, ChromaDBVectorMemory) but require explicit implementation.

Our approach: Implement memory management from the start. Never assume context will stay within limits at scale.

Non-Determinism at Scale

LLM variance compounds across multiple agent exchanges. Same scenario, different results every run. Debugging requires tracing emergent behaviors across agent interactions - traditional stack traces are insufficient.

Our approach: Structured outputs at every agent boundary. Diverse test sets. Accept variance in phrasing; test for structural correctness and factual accuracy instead.

Hard Iteration Limits

Every agent that can loop needs a maximum iteration count. The single most important safeguard against cost spirals. Not optional.

Cost Circuit Breakers

Alert at 5x and 10x expected cost per operation. Automatic halt on anomalies. Build these in from sprint one, not after your first billing surprise.

Structured Outputs

Define Pydantic schemas for inter-agent messages. Makes handoffs explicit, testable, and reliable. Reduces hallucination propagation between agents significantly.

HITL for Consequential Actions

Any workflow that writes to external systems needs a human-in-the-loop checkpoint before the write. Agent mistakes in multi-step systems are difficult to reverse.