A principle for agentic work
Finish lines, not managers.
The Relay Principle is a model for long-horizon agentic work without hierarchical control.
Define the goal. Define the edges. Then step back.
The Relay Principle
Finish line over plan
A plan is valid until reality diverges from it — which it always does. A finish line stays valid. The runner adapts the path; the goal does not move.
Scope over script
The boundary is what enables the trust. Define the runway clearly — then get out of the way. A scope is an invitation. A script is a leash.
Handoff over handhold
Each runner completes one leg, logs what they found, and passes the baton — cleanly and completely. No supervisor relays messages between them.
Record over memory
Shared knowledge lives in documents, not in any one agent's context. What travels down the chain is the record of work — not the worker who did it.
Signal over spin
A runner that stops and tells the truth about being stuck is more valuable than one that invents a confident-looking next step. Clarity is the handoff.
The principles
These are not rules for the runner. They are obligations of the designer.
- iDefine a finish line. A relay without a concrete end state runs forever. The goal must be achievable and recognisable.
- iiSize one leg. Not so small it is trivial, not so large that context bloats and containment breaks.
- iiiState the handover explicitly. Every runner must know how to pass to the next — what to write, where to put it, how to spawn.
- ivDefine the intervention signal. Decide in advance what should make a runner stop and call for the human.
- vTrust is earned per leg. Every runner re-anchors to the goal. Drift is a signal, not an excuse.
- viLetting go is the point. The next runner is trusted. The log is the thread between sessions.
Why?
Most AI efforts fail quietly for one basic reason: we try to make AI behave like people.
We wrap it in roles, orchestrators, and hierarchies. We assign responsibilities. We design handoffs. It looks structured, familiar, and safe.
It is also wrong.
Those patterns come from human limitations: communication overhead, slow learning, specialization, bias, and misalignment. They are solutions to problems AI does not have.
AI does not need coordination in the way humans do. It does not get tired, territorial, or stuck in its lane.
Yet we keep rebuilding the same structures. We design layers, roles, and boundaries not because AI needs them, but because we do.
That is the trap.
The real opportunity is not to fit AI into existing workflows, but to question the assumptions those workflows were built on.
What does a workflow look like when those human constraints disappear?
Fewer boundaries. Less hierarchy. More fluid execution.
This is the shift in ways of working: not AI inside old systems, but entirely new systems shaped around what AI actually is.