Think about how almost everyone uses AI right now. You open a conversation, do one job, close it. Then you open another one for the next job, paste in whatever the first one produced, and fix the parts that got summarised away on the way over.
Every tool is a single instrument, and you're playing all of them, in sequence, by hand.
That works up to a point — further than you'd expect, which is why so many businesses settle there and believe they've finished. It stops working the moment the work stops being one job and becomes a relay.
Real work isn't a job. It's a relay.
Nobody's actual work is one task. A campaign is a strategy, which becomes copy, which becomes assets, which becomes placement, which becomes measurement, which becomes a change to the strategy. That's five handoffs at minimum, and each one depends on the last.
In a human team, every handoff loses something. The strategist's intent gets squeezed into a brief. The brief gets read by a designer who wasn't in the room. The designer's choices get flattened by whoever places the media. Everyone has lived through a project where the thing that came out was technically what was asked for and clearly not what was meant.
Faster machines don't fix this
The instinct is to assume handing work between software agents will be cleaner than handing it between people — software doesn't forget, doesn't editorialise, doesn't have a meeting to get to.
The truth is closer to the opposite, for one reason. A person who drops something usually feels that something is missing, and asks. A machine that drops something doesn't notice, doesn't ask, and hands you a confident, well-formed output that's missing a constraint nobody will check until the work is live.
Speed doesn't remove the loss. It removes the friction that used to make the loss visible.
The handoff contract
That turns the complaint into a rule you can design against.
The handoff contract is exactly what passes between two stages: what gets transferred, in what form, checked against what, and remembered where. It's the thing that decides whether a relay of agents is a system or a run of plausible accidents.
A contract isn't a message. Asking one agent to write a summary for the next one is a message. A contract says the summary must carry the audience, the constraint, the decision that was already made and the reason it was made — and that the next stage isn't allowed to start if any of those are missing.
Four things every contract has to nail
- The payload
Exactly which facts, decisions and constraints move to the next stage, in a shape both stages understand.
- The standard
What the next stage checks the payload against before it starts — so a missing piece gets caught at the seam, not six steps later.
- The authority
Which calls the next stage can make on its own, and which have to go back to whoever owned the intent.
- The record
Where the handoff is written down, so what got passed and what got dropped has an answer that doesn't depend on anyone remembering it.
The agents are the easy part. What passes between them, gets checked, and gets remembered — that's the engineering problem.
The three ways it breaks
Every failure in a relay of agents comes down to one of three things, and they're easy to miss, so learn their names.
- Lost context — a constraint from three stages back never made it forward, and nothing flagged it.
- No shared memory — every stage starts from zero and re-decides things that were already settled.
- Nobody owns the outcome — every stage did its job correctly, and the finished thing is still wrong.
All three can happen to a team of brilliant agents. That's the point. How smart the parts are isn't what saves you. The seams are.
What this means for you
If you're running AI tools one at a time and connecting them by copying and pasting, you haven't built a system. You've built a relay where you are every handoff — and you're the one quietly dropping things, because you're human and it's the end of a long day.
The fix is deciding once what has to pass between the stages and what has to get checked — and then holding every step to it. That's a smaller job than it sounds. And it's the difference between a system and a series of accidents that happen to look like work.
