Microsoft Conductor

The deciding trait is that the whole pipeline is one file you commit. Agents, prompts and the wiring between them sit in version control, so a change to how work flows is a diff somebody approves rather than a conversation you half remember. That alone rules out a whole category of Monday-morning confusion.

What you give up is improvisation. If your work does not decompose into named steps, writing it down is friction rather than clarity. Pick it when the process is stable and worth freezing. Pick Schaltwerk if you would rather steer sessions by hand.

The failure mode is ordering. Conditions are evaluated top down and the first true one takes the run, so a broad condition placed above a narrow one swallows every case beneath it. Nothing warns you, and the run finishes looking exactly like a correct one. Debugging that means reading the file and simulating the expressions in your head.

Two guards bound the blast radius, and the documentation names both: a maximum iteration count and a timeout. Those stop a runaway. They do not stop a wrong answer, which is the more common outcome.

1. Separating orchestration from inference is the principled decision here. A routing choice made by an expression is reproducible and inspectable; one made by a model is neither, and the difference compounds across long pipelines. 2. Parallelism is declared in two documented shapes, static groups and for-each fan-out, so concurrency is a property of the definition rather than an emergent behaviour.

2. No evaluation accompanies any of this, and none is owed. The project asserts a structure, not a capability, and a structural argument is assessed by reading it.

Ownership is the whole analysis. There is no cap table, no runway and no acquirer, because the acquirer already published it. What replaces funding risk is portfolio risk: large vendors keep many small repositories and retire the ones that do not feed a product line they already sell.

Moat: distribution through the parent's developer estate, obtained at no cost to the project. Likely path: quiet absorption into a larger tooling line, or a slow archive with the source left standing. Position: build on the file format, not on the maintainer.

Procurement is trivial and governance is absent, which is the trade. The licence costs nothing at any headcount, and the only new line item is inference the agents consume. Exit codes mean a workflow becomes a pipeline step our existing runners can gate on, which is the part I actually wanted from this category.

Against that: no single sign-on, no user directory, no record of who changed a definition beyond what repository history shows. Approved with conditions: definitions live in a protected repository and changes take the same review as production code.

The licence is MIT and the artefact is text, which is most of what I ask for. Prompts are not buried in a binary; they sit beside the routing in one document, so tuning an agent is an edit and a commit. Providers are mine to choose per step, a Copilot login or an Anthropic key.

What I cannot do is attach the servers I already run, because this speaks neither side of MCP. That is a real hole for anyone whose tooling lives behind that protocol, and it is the first thing I would fork to fix.