Opinions backed by things I actually shipped.
Different project, different language, 17 days apart. The mechanism is boring on purpose: a failure written down as a checklist question outlives the fix, and the next reviewer, human or agent, has to answer it.
The bottleneck in working with agents was never the decision. It was being the messenger between two machines. Here is what happened when I removed the relay and kept the decision, including the moment my own approval got thrown out.
One person plus agents is a real operating model in 2026. The design question is not what the agent can do, but which verbs you refuse to delegate: push, merge, declare done.
A portability stance, not a privacy one: keep your project's history, decisions, and state in three plain files you own, and any AI tool can pick up exactly where the last one left off.
A practical follow-up: compacting, context trimming, handover files, and model routing to reduce token overhead in long agentic sessions.
In my agentic AI sessions, 85-98% of token spend was overhead. Here are the changes that actually reduced it.