CrewAI, the Multi-Agent Framework I Tried Three Times

CrewAI is the framework I keep almost adopting. The conceptual model is the cleanest I have found in the multi-agent space: define agents with roles, give them tools, give them a task, and let a manager agent orchestrate. It maps to how I actually think about empire workflows.

I tried CrewAI for the autokaam social composer in 2025. Three agents: scout (find a recent article), critic (verify the facts), composer (write the social post). It worked on the first try. It also took 4x more tokens than my hand-written equivalent and added 3 seconds of latency, because the manager agent had to do a planning round before each task.

I tried it again for a recon pipeline. Same outcome. Tasks completed, beautifully orchestrated, twice the cost.

The third attempt was the kaam-tracker invoice-collection flow. CrewAI agents are call-out-and-collect-info, write-the-message, send-it. Here CrewAI shone, because the human-in-the-loop pattern actually fits the framework's strengths. I shipped it. It runs every Tuesday morning. It costs more per run than a hand-rolled script, but the maintenance is lower because the abstraction is correct for the problem shape.

My CrewAI rule is: if your flow looks like "team of specialists collaborating on a task with coordination", CrewAI is genuinely the right abstraction. If your flow looks like "do these five things in sequence with retries", it is overkill.

For 2026, the multi-agent space is consolidating around a few patterns: CrewAI for explicit-role flows, LangGraph for state-machine flows, AutoGen for negotiation-heavy flows, plus a long tail of vendor-specific kits like Anthropic Skills and OpenAI Swarm. I would not adopt any of them without a real workload to test against.

The Indian-operator angle is cost again. Multi-agent flows multiply token spend. Test on Cerebras-Qwen or Gemini Flash before you scale up to Sonnet. The CrewAI samples in the docs all use GPT-4 class models, and for most tasks that is genuine money you are leaving on the table.

I respect the project. I just use it sparingly.

Agentic workflows are the dominant production shape in 2026. CrewAI is one of three frameworks (alongside LangGraph and AutoGen) that have survived the 2025 churn. The conceptual model maps cleanly to specialist-team workflows, but token cost and latency overhead make it unsuitable for simple sequential flows.