- Six names, six answers to one question
- The feature table is the wrong scorecard
- Operability, scored in plain words
- The pick rule
Six frameworks now coordinate most production agent systems. Each one answers the same question differently. Who decides what runs next, the code or the model? The shortlist is current as of this week.
One entry breaks the pattern completely. It has no orchestration layer at all. That detail matters more than any feature row. It waits in the list below, in its place.
The short answer first. LangGraph gives developers an explicit state graph. CrewAI gives agents roles, goals and a process. Microsoft Agent Framework targets organizational workflows with checkpoints.
OpenAI Agents SDK lets the agent decide handoffs. Google ADK splits work from a root agent downward. Claude Agent SDK keeps orchestration inside the agent. Six designs, six different debugging stories at 2 a.m.
Six names, six answers to one question
LangGraph models a workflow as nodes and edges. Shared state moves along those edges between steps. Its own documentation calls it a low-level orchestration framework. It persists state and supports human-in-the-loop pauses.
Read the shape yourself in the LangGraph overview in its documentation. CrewAI starts from the org chart instead. Every agent carries a role, a goal, a backstory. Crews run tasks sequentially or under a manager agent.
Microsoft Agent Framework aims at organizational workflows specifically. It also uses graphs, with branching built in. Checkpoints and human intervention pauses come standard there. OpenAI Agents SDK takes the runtime approach.
The agent itself decides whether to delegate control. It can also keep execution and finish alone. Google ADK works as a hybrid between them. A developer defines the strategy for running tasks.
A root agent then splits work among sub-agents. Claude Agent SDK removes the separate layer entirely. Orchestration lives inside the agent loop itself. The agent gathers context, works tools, evaluates, continues. It stops only when no tool calls remain.
The feature table is the wrong scorecard
Feature tables reward the longest checklist. Production punishes the hardest debugging story instead. Your agents will fail on a holiday weekend. The question is who can read the failure.
OpenAI LifeVal grades AI agents on errands that stop short of the CRA login, and grading matters. But a benchmark score never restarted a stuck run. State visibility decides the first hour of every incident. Notion agent Memories are private to each user. Decide what they may keep raises the same question.
What the system remembers, someone must be able to inspect. Tooling trust carries the same weight as design. The malicious Claude Code extension lasted 11 hours. A one-day install delay would have skipped it, which was an operations failure. Ownership is the last line of the scorecard. Ask who signs Ada agent refunds before the Dialpad close names an owner per action. A framework without an owner per run drifts.
Operability, scored in plain words
Score each candidate on four plain questions. Can your team read the state mid-run? Can a human pause it before damage? Can a failed run restart from a checkpoint?
Can a new hire explain yesterday’s bad run? LangGraph answers through its graph and persistence. Microsoft answers through checkpoints and intervention pauses. CrewAI answers through roles a manager can inspect.
The SDK designs answer through the agent’s own trace. None of these answers is free, though. Each moves complexity to a different desk. A graph makes the flow visible and demands modelling upfront.
A role system starts fast and hides its wiring. A runtime agent decides quickly and explains slowly. State design is the hidden line item everywhere. Shared memory and context pass between every step.
Whoever owns that design owns the incident queue. Budget for that ownership before the first deployment, not after. Alberta product teams are choosing stacks this quarter. Their on-call reality should outrank any demo. The best demo rarely survives contact with payroll data. Pick for the worst night, not the best morning.
The pick rule
Shortlist two frameworks, never six, before building. Frameworks will keep merging and renaming themselves. The operability questions will still be the same four. Ask them in the same order every time.
Write the 2 a.m. test down for both. Who restarts it when a run stalls? Where does its state live between steps? How does a human pause it mid-task?
The framework with the shorter answers wins. Send those answers to your tech lead in writing. One honest note before you run the test. No hands-on framework test ran for this piece. The test above is yours to run, deliberately.
AI Magazine Canada tests the tools Canadian teams actually buy, one plain verdict at a time, and the next verdict lands tomorrow morning.
