AgentField

The trait that decides it is that an agent is a function. You write ordinary code in a language you already use, and the platform makes it an endpoint, which means your existing testing, logging and deployment habits all still apply. Nothing new has to be learned before something useful exists.

What that costs you is the visual overview some teams want, because everything is code. Pick it if your instinct is to write a function. Pick Dify if your instinct is to draw a diagram.

The headline capability is the hazard. Routing, queuing and retries all live inside the control plane, which is convenient until a retry policy meets a recursive call graph, at which point a single request multiplies into an invoice. No documented concurrency ceiling, no depth limit and no circuit breaker appear in the material.

What it does right is visibility. Tracing is part of the control plane rather than an integration, so when the fan-out does misbehave the record of it already exists.

1. Treating coding harnesses as interchangeable behind a single call is an elegant abstraction and a strong claim, because those harnesses differ in context strategy, edit format and verification behaviour. 2. Swapping one for another should therefore change results in ways the interface hides.

2. No comparison is published: no task set run across the six options, no success rates, no cost differences. The abstraction is offered on the assumption that the choice does not matter, which is exactly the assumption worth testing.

Control planes are infrastructure, and infrastructure monetises through operation rather than through code, so the absence of any hosted offering leaves the value entirely uncaptured. That is fine while the goal is adoption and unsustainable once maintenance becomes a job.

Moat: none yet; the harness abstraction is copyable in a weekend by anyone who reads it. Likely path: a managed control plane appears, or a larger platform absorbs the pattern. Position: run it yourself, keep your agent functions plain, and make sure lifting them out later is a rename.

The useful half is that work runs without a person present, which means it can be scheduled, measured and attributed, and the emitted deployment stack gives my platform team something familiar to operate rather than a bespoke service.

The awkward half is the install path, which registers a background service and a menu-bar item on developer machines. That is a desktop tool wearing infrastructure clothes, and I would rather one managed deployment than sixty. Support is a repository. Approved with conditions: a central deployment only, and no installer on individual workstations.

Serving the protocol as well as speaking it is the part I care about, because it means a coding agent I already use can drive this thing rather than only being driven by it. That inversion is rare and it is the difference between a framework and a component.

The routing library in the middle means model choice is a string, including an endpoint on my own machine, so nothing forces a hosted provider. Permissive licence, three languages, readable install. This is a genuinely well-shaped piece of plumbing.