Laddr

The deciding trait is that a coordinator agent does the routing for you. You describe specialists, hand the coordinator a task, and it decides who does what and stitches the answers back together, which is the part most people write badly by hand. If you have ever built a router out of if-statements over intent labels, you know why that matters.

You will want something else if your problem is one agent with good tools, because the delegation machinery buys you nothing there. Pick it for genuinely parallel work across several specialists. Pick Temporal if what you actually need is durable workflows.

The box is emptier than the pitch suggests. Agents call tools, and every one of those tools is Python you write yourself: no shell, no file editing, no git. The framework moves messages between things you have not built yet, so the distance from pip install to a system that does work is measured in your afternoons, not in config.

That is defensible for a library, and it is the wrong expectation to arrive with. What it does right is failure containment. A worker dying leaves the rest of the system running, which is more than most orchestration code manages on its first outage.

1. Every agent action is recorded to SQLite or PostgreSQL, with optional Langfuse spans, which makes the execution trace a first-class artefact rather than a log line. For a delegating system this is the only way to answer which component produced a given output, and most frameworks in this category leave it to the user. 2. The deterministic mode declares inputs, outputs and dependencies per step, so a run is reproducible in the sense that matters: the same graph, not merely the same prompt.

2. No evaluation is published. Nothing here claims a capability that would require one.

Agnet Labs has shipped a permissively licensed framework with no hosted tier, which is the standard opening move and leaves the monetisation question entirely open. Three hundred and forty-three stars is a seed-stage signal: enough to prove somebody cares, nowhere near enough to price anything.

Moat: none yet. The plausible one is the managed control plane, because self-hosting a queue-backed agent fleet is exactly the pain a paid tier sells against. Likely acquirer: a platform vendor that wants a multi-agent runtime it did not have to write. Position: fine as a dependency, too early to bet a roadmap on.

Zero licence cost for sixty engineers, and then the infrastructure bill arrives. This wants Redis and a database running before anything happens, plus a FastAPI service to expose it, which means my platform team owns three more things in the on-call rota. That is not a purchase decision, it is a staffing one.

There is a metrics dashboard, which is the only operator surface offered: no seats to provision, no directory integration, no retention policy to hand the security questionnaire. Nothing here is triggered on a schedule either. Not yet, and not as a product; it may be a dependency inside a service we already run.

Apache-2.0 with no clauses to argue about, and the model layer is a plain import: provider helpers live under laddr.llms and the generated starter hands you gemini("gemini-2.0-flash") to replace. Swapping that is an edit, not a config schema I have to learn, which is the right shape for a library.

What I do not get is MCP. Every server I already run has to be rewrapped as a Python tool before an agent here can touch it, and that is real work I have done once already elsewhere. The licence means a fork stays viable, so I can add the client myself if it matters enough.