A dependency is a promise that another team must keep before yours can finish. Most 'collaboration issues' are unmanaged dependencies wearing a smile. Visualise them early, reduce them on purpose, and give the remainder an owner who will notice when the promise slips.
See them before they bite
Write dependencies as team A needs X from team B by date, or this feature cannot meet its PI objective. 'We should stay in touch' is not a dependency. It is a social hope. Put the real ones on a program board, a dependency board, or a simple table if your tool is theatre: from, to, what, needed-by, status. If it is not visible to both teams and the RTE, it does not exist until it explodes.
Hunt the types, not just the tickets. Sequencing dependencies (we cannot test until you ship the API). Shared-people dependencies (the only security reviewer is on three trains). Shared-system dependencies (one environment, six teams, a freeze nobody scheduled). Knowledge dependencies (only one person understands the billing engine). People-dependencies are the ones organisations refuse to write down because the org chart finds them embarrassing. Write them down anyway.
Do this in PI Planning and again every week in the ART sync. A dependency discovered in the last iteration of the PI is an issue with extra paperwork. Walk the board and ask, for each red string: is the providing team still on the hook, and has anything moved? Silence is how strings become folklore.
Reduce first. Manage what is left.
The cheapest dependency is the one you delete. Split the feature so one team owns an end-to-end slice. Add a temporary mock or feature flag so the consumer is not blocked on a perfect provider. Move a specialist into the team for a PI instead of booking them as a visiting deity. Architects who keep drawing more integration points are not 'aligning the enterprise.' They are manufacturing queueing theory.
What you cannot delete needs a contract lighter than a legal one: an agreed interface, an integration date, and a fallback. 'We'll pair in Slack' is not a contract. 'Contract tests in the shared repo by 12 May; if they fail, consumer uses the stub and the feature is uncommitted' is a contract. Put the fallback on the board. A fallback invented during the outage is just adrenaline with a Jira key.
Cap work-in-progress that is waiting on another team. A board full of blocked cards is not a busy team. It is a queue you are too polite to refuse. Pulling more local work while a critical dependency sits amber is how you hit sprint goals and miss the increment. That is a local win and a train loss. Say it in the room while there is still time to swap scope.
Own the handshake
Every material dependency has two owners: a provider who can actually do the work, and a consumer who will escalate when the date moves. 'The platform team' is not an owner. Neither is 'architecture.' If the only person who can open the firewall is in another directorate, that person is on the board whether their manager enjoys the visibility or not.
Integrate continuously, not in a ceremonial week-twelve 'merge fest.' Small, frequent integration makes dependency slippage a this-week problem. A single integration weekend at the end of the PI is how you learn — loudly — that three teams built compatible stories and an incompatible system.
Report dependency health the way you report risks: new ones, slipped dates, and decisions needed. Leadership does not need every string. They need 'Feature 14 will miss unless Team B's API lands this Friday or we descope the mobile client.' That sentence is why the board exists. A wall of red wool with no ask is interior decoration.
Summary & next steps
Write dependencies as who, what, and when. Visualise them where both teams can see them. Delete what you can; contract and integrate the rest. Name owners, not departments.
Next: portfolio kanban — how work above the train is supposed to flow, and why an epic parking lot is not a strategy.