what is the best multi-agent system?

build or run, decide that first

Frameworks give you primitives and leave the operations to you. Langgraph for explicit state and resumable runs, crewai for roles and tasks working today, autogen for conversational group patterns, the openai agents sdk for handoffs with tracing. All free, all capable, all leaving the same gap.

Platforms give you a place agents run, with the operations included. You trade some control for not spending a quarter on permissions, memory, and audit.

Anyone comparing a framework against a platform on capability is comparing an engine against a car.

the criteria that actually separate them

Not the model. Every serious option can use the same models. Five things:

1. Where shared context lives. If agents pass summaries to each other, context is lost at every handoff. A shared store that all agents read is the difference between three agents working together and three agents guessing.
2. Whether the reviewer is a separate agent. Self-review checks against the assumptions that produced the error. This single decision moves output quality more than any prompt work.
3. Permission scoping. Per agent, per project, revocable, without a support ticket.
4. Isolation. Whether one project's data can reach another's context. This is where most homegrown systems quietly fail.
5. A stop. A global step budget, a spend cap, and a human gate. Multi-agent systems fail toward loops between agents, which per-agent limits do not catch.

the sizing rule nobody publishes

More agents is not better. Coordination overhead grows faster than the benefit, and past roughly five agents in a single run you are usually paying for confusion. If a design calls for twelve, the real problem is almost always three agents with badly designed tools.

Similarly, do not start multi-agent. Run one agent on twenty real tasks, write down exactly where it failed, and add a second agent only if the failure was one of the two things a second agent fixes: too many tools to choose between, or bad self-review.

the cost side

Several agents means several context windows, and each handoff carries history. A three-agent run typically costs two to four times a single-agent run on the same task. That is worth it when quality matters and verification is expensive, and it is waste on a well-specified single task.

how this works in aldena

Aldena is a platform built directly against those five criteria. Shared context is room memory, 100 entries every agent reads plus 50 private each, so a handoff carries facts rather than a summary. The reviewer is a separate role, one of eleven, placed above the workers in an org chart.

Permissions are per agent and per room, isolation is physical because every room has its own server and credentials, and the stop is an approval gate on by default before anything irreversible. And you watch the coordination happen live rather than reconstructing it from a trace.