Watch someone delegate to an AI agent for the first time. They hand it one task and wait. They review the result, then hand it the next. When they add a second agent, they worry the two will collide, so the second waits its turn. Within days they have rebuilt, without noticing, the working conditions of a junior employee: one assignment at a time, attention during office hours, coordination by queue.
A manager handing out one assignment at a time is often arranging work around their own attention. Office hours fit human schedules; a seat fits a human body. Meetings help people exchange context and make decisions together. Applying these habits by default to agents can turn independent tasks into a queue.
I hold my own system to this principle: derive each agent's working pattern from the constraints of the work, and keep human pacing where judgment or relationships require it.
Several sessions, one repository
A shared codebase needs protection against conflicting edits and lost context. Two writers must not silently overwrite each other, and someone must own the canonical record of what happened. Independent tasks can proceed together when those obligations are met.
My agent sessions run in parallel, sometimes several on the same repository. Each works in an isolated checkout and registers the areas it intends to change on a shared board outside the repositories. The protocol requires sessions to read the board before writing and at checkpoints.
The claims are advisory locks: declared and visible rather than enforced by the filesystem. Every session is required to honor them. When two claims collide, the colliding write waits and everything else proceeds. Sessions leave messages on the board for others to read at their next checkpoint. One session owns each project's canonical context; the others write their results to separate artifacts the owner reconciles.
The protocol is public in my synthesis-skills repository. It uses ordinary files and isolated checkouts to preserve ownership, share context and let independent work continue.
A session can read the shared record and continue when its dependencies are ready, without waiting for everyone to gather.
The whole record, a day early
A person who starts preparing shortly before a meeting often has time only to skim. Reading the full history takes longer than the calendar allows, so they surface whatever is most recent and hope it is also most relevant. I got a clean demonstration of the cost when a high-stakes weekly was rescheduled onto the same day and the prep, requested sixteen minutes before the start, led with my loudest recent workstream instead of the mandate the person across the table actually held. Recency bias with a deadline.
For meetings I flag as high-stakes, my system prepares a briefing one business day ahead, built from the durable record: the full transcript of the previous meeting, the counterpart's priorities from the documents where they stated them, the project notes that name them, and open commitments in both directions. The briefing lists which sources it read and the date of the newest one, so I can tell a researched pack from a summary of recent activity. The lead time pays for that reading and leaves me time to decide what the meeting needs.
Whose clock it is
My mornings start with work that finished before I sat down. The evening close generates the next day's meeting briefings. When the agents run on a schedule, the daily brief is ready before my review window opens instead of being assembled while I watch. When I step away in the middle of the day, sessions keep working and play a chime when something completes or a decision blocks them. The chime says nothing more than that, because speakers get overheard.
Messages to people still run on a human clock. Recognition loses value overnight, so unsent kudos carry expiry dates, and the evening close forces a send-or-release decision on each one. Morning updates land in the morning, when the humans reading them are awake and reading.
The gates that stay human
Nothing ships in my voice without me. Drafts queue for one-tap review: approve, edit or skip. I retain decisions about opinions, commitments and sensitive relationships. The narrow class of messages I permit agents to send (scheduling pings, receipts, follow-up nudges, information relays with citations) must be labeled as agent-sent, bounded by a matrix I wrote in advance, and logged where I will read every send. Each proposed send still needs my approval. A diverged Git branch comes to me for a decision. Production deployments and anything public-facing wait for my explicit go each time; yesterday's approval does not carry forward.
These gates protect decisions that are expensive to reverse and outputs that carry my name and my relationships. My judgment is the product at that point in the flow. The agent should bring the diff, the thread, the evidence and a recommendation, making the decision better and cheaper. I want to spend seconds making a well-prepared decision instead of an afternoon reconstructing what happened. Automating the work and automating the judgment are different projects. Confusing them is how you end up with a system that signs your name while you sleep.
A test you can run
Take any rule in how your agents work today: when they run, how many run at once, how they coordinate, how much context they read before acting, what they may do without you. For each rule, ask which constraint it serves.
It serves a constraint of the work: two writers on the same files, an external rate limit, an ordering some dependency imposes. Keep the rule, and write the constraint down so the next person knows why it exists.
It serves a human limitation that arrived by habit: serial attention, office hours, coordination by meeting, preparation by skim. Remove it, then look at what the work becomes without it. Independent tasks can finish together; the full record can be read before you arrive.
It marks a judgment gate: an irreversible action, an external commitment, your name on the output. Keep it human-paced on purpose, and spend your engineering on feeding it well.
I run agents because the work deserves better than my limitations, and I keep the gates because parts of the work are my judgment. The org chart was always a compromise with biology. The judgment gate never was.