Working agreementMaintained by Soph
Members · 5
thehumanailoop.com/methodology
This is a working simulation of a sprint planning session run by a mixed human and AI team inside a shared workspace. It answers a specific question: what does it actually look like when AI is a teammate rather than a tool?
Two different companies' models are doing four different jobs on one team. That is deliberate. The AI members hold different mandates, and one of them is set up to widen the option space while another narrows it. Assigning every job to a single AI collapses the distinction and produces confident, average answers, neither wide enough to surprise you nor sharp enough to ship.
Nearly every AI workflow today requires the human to state the requirement completely upfront. That is the waterfall problem with the signal turned off, and it fails for the same reason long specifications always failed: one person cannot think of everything before the work starts.
Watch the Working Agreement field instead. It opens with one line and closes with four or five, depending on the path you take. Every entry is attributed to the human, because every entry required something only she had. Nobody wrote that document in advance. It was complete when the work was.
Twice in the session, an AI member stops because it has hit something no amount of processing will resolve. Once when two openings serve two different audiences and only the human knows which room this is for. Once when the team sizes the same work and disagrees wildly, which turns out to mean the brief said different things to different readers.
That second one borrows a ceremony the industry already invented. In planning poker the number does not matter and the spread does. Wide disagreement means people are imagining different work, which means somebody is holding a requirement that never made it into words. Agile built a detector for unspecified intent twenty years before anyone needed it for this.
Each time, the machine states what it cannot determine, why it cannot, and what is blocked. The human answers in one line. The answer becomes a locked decision with attribution, and the work resumes.
"Direction" gets treated as one thing a person either supplies or does not. In practice it is four distinct moves, and a system that supports only some of them is a tool rather than a teammate.
Three of these have obvious expressions in existing products. Suppress has almost none. It is also the move that most clearly separates a collaborator from a tool, because it is the only one where the machine is right and gets overruled anyway, on knowledge it structurally cannot have.
Three of the four appear on any path through the session above. The Sidebar prep exchange is Suppress: the recommendation is correct on the data and wrong in the room. What matters is not that it gets overruled, it is that the reason gets recorded, so the next session does not relitigate it.
The fourth, Steer, is the re-loop. You only see it if you take that path.
A dedicated #triad-state channel carries the team's working memory as structured fields: the working agreement, sprint status, active priorities, locked decisions, pinned artifacts, and current handoff. Any member can read the team's state without rereading the conversation.
This is the piece most human-AI workflows are missing. Context lives in someone's head or scattered across threads, so every session restarts from zero and the same decisions get relitigated a week later.
Note who maintains it. In a human team, keeping the record current is a separate chore that somebody does late and inconsistently, which is why coordination is expensive enough to need its own role. Here the member who synthesizes the discussion is the member who keeps the record, so the record is a byproduct of the thinking rather than a task anyone has to remember.
Work moves between members on stated conditions rather than assumption. Decisions are locked with attribution, and the locked list is the source of truth.
Partway through, the simulation branches and the human chooses: lock the plan and begin execution, or re-loop to refine before committing. The re-loop path is the more interesting one. It shows targeted revision that preserves everything already decided, a scoped pass on one section with a re-loop count of one and context loss of zero. It also produces an additional line of specification that would otherwise never have been written down. Most AI workflows treat revision as a restart, which is why people stop revising and ship the first plausible draft.
Human-AI collaboration fails at the seams: unclear ownership, no shared state, no record of what was decided, and no distinction between an AI that should widen your thinking and one that should close it down. Roles, shared state, explicit handoffs, and a requirement that accretes are what make a mixed team behave like a team rather than one person with several chatbots open.
This prototype is an applied demonstration of The Human-AI Loop methodology and the Triad operating model. Related work: The Loop Simulator · The Strategist as Conductor · The AI Literacy Library