agentUniverse

You will save a week here if your problem matches one of the shipped patterns, because the hard part of multi-agent work is deciding who checks whom, and that decision has already been made and named for you. The deciding daily trait is that a review step is built into the shape rather than something you remember to add at the end when the output looks wrong.

Pick it if your task decomposes into planning, doing and checking. Pick MetaGPT if you want a role-based crew with a larger library of examples to copy from.

The risk is fit. What is being sold is a factory of collaboration shapes, and a factory is only useful if one of its outputs matches your problem. When none does, you are back to writing an orchestration loop by hand with a framework's abstractions in the way. There is a second cost hiding in the design: an iterating review stage means every pass through the work happens at least twice, in tokens.

What it does right: the shapes are declared in configuration, so switching between them is an edit rather than a rewrite.

1. PEER assigns planning, execution, expression and review to distinct agents and iterates on the review's feedback, which puts verification inside the loop rather than after it. That is the correct place for it. 2. DOE targets data-heavy tasks needing expert judgement by chaining refinement, opinion injection and expression, a narrower and more opinionated shape.

Neither pattern is accompanied by a measurement, so the claim that iteration improves output is asserted rather than demonstrated. The observation: a framework that names its review agent has already thought harder about correctness than most on this board.

The provenance is the strength and the constraint. This came out of a payments business at scale, so the patterns were built against real workloads rather than a demo, and a company that size does not abandon infrastructure it depends on internally. It also means nobody outside Ant is a customer, there is no price, and the release cadence follows an internal roadmap you cannot see or influence.

Likely path: continued maintenance as a standards and recruiting exercise. Position: a safe dependency, a poor partnership, and no leverage if a pattern you need is missing.

The demo is a plan and a review agent arguing productively. As a dependency it costs nothing to install across the team, and the spend lands on whichever provider is wired in through configuration, which for us is a contract we already hold. There is no identity story and no audit surface, because a Python library has neither, so both become the responsibility of whatever service wraps it. Nothing here runs in a pipeline on its own.

Onboarding is three days, mostly learning the pattern vocabulary. Approved with conditions: one team, our keys, our wrapper.

Both directions of the protocol is the detail that matters. Most frameworks consume tool servers; this one also exposes itself as a server, so an agency I build here becomes a tool something else can call, and I stop writing glue between two ecosystems. Apache-2.0 makes that permanent, and one pip command installs it.

Model wiring is configuration across Qwen, DeepSeek, Gemini, Llama and Kimi, so switching vendors is a file edit. What is missing is a documented local runtime, so my own box is reachable only by pretending to be somebody's API.