Dani Shemesh

Whether to move a development procedure out of the agent session: what actually justifies it, what it costs to live with, and the one risk no engineering fixes.

One workflow engine read from the inside: the four phases a run moves through, and the bookkeeping decisions that separate working from mostly working.

What using a workflow engine is like: the front doors, the reusable pieces, the five claims Archon makes for itself, and the rough edges that come with it.

What separates a workflow engine from an orchestrator skill, what that one inversion buys, and how little of what is sold as an engine actually qualifies.

Six limits an orchestrator skill cannot write its way out of, because they follow from where the procedure runs rather than from how well it was written.

Honest limitations of agentic development workflows, all seen in real runs, and the short list of what to build first if you are starting one.

Packaging a workflow so others can depend on it, porting it to another harness, the road to unattended runs, and where the whole design still fails.

Three axes of evidence for an agentic workflow: is the output correct, where does the process hurt, did anything ship. And how each one gets gamed.

Item size, what a gate really costs in human attention, which decisions deserve the expensive model, and what the whole apparatus buys you.

The principles that keep an agentic workflow honest: where run state lives, what makes a gate a real check, and who gets to decide when a build loop stops.

What an agentic development workflow is: a phased procedure that makes a model's work inspectable and interruptible, and the parts it is built from.