In brief: No role can delegate more work to agents than the Organizer — routing, chasing, status-gathering, preparing — and no role loses more when that delegation goes quiet, because coordination is invisible when it works. A human coordinator who stops is noticed by lunchtime; an agent that stops chasing is noticed at the deadline. This briefing is for programme and project leads, and for the leaders who depend on them: it covers what to delegate, what to keep (sequencing, priority conflict, relationship repair), how to start when IT will not connect anything, and the line that must be drawn before any of it — coordination agents report on work items, never on people.
No role has more to delegate to agents than the Organizer — and no role loses more when the delegation goes quiet. This briefing follows one cross-functional project through an Organizer’s hands: what agents can chase, watch and prepare, what only she can sequence, and how coordination stays visible while machines do the legwork.
On Monday morning, Maren Voss inherits a yes. At Bruns & Lehmann, the family-owned pump manufacturer this Path follows — 420 people in northern Germany, customers across the continent, an ERP older than most of the vendors now pitching to replace parts of it — the advisory board approved the automation of supplier-order intake on Friday. That was the decision the previous briefing in this Path followed through the Owner’s hands. Maren is the programme lead, in practice the person everyone calls when two departments disagree, and the delivery is now hers. It is a textbook coordination problem: an external integrator, the IT team touching a part of the ERP nobody has touched since 2019, operations who want it yesterday, quality assurance who want it tested twice, and a go-live date that the sales director has already mentioned to a key account.
Maren did not wait for this project to start working with agents. A year ago she built herself a meeting-preparation assistant, because preparing steering meetings was eating her Fridays: it pulls the open items from the project tool, drafts the agenda from last week’s minutes, and lists what changed. Nobody assigned it to her. She started on her own pain, IT later gave it a proper identity, and it has quietly become how she runs every project since. That order — the person who lives the coordination designs the help, engineering hardens it afterwards — is the pattern this whole briefing assumes.
What an Organizer actually holds
The Compass groups the delivery roles of a transformation as Transformers — people, not the neural-network architecture; the difference is exactly the accountability gap this series works. The Organizer is the Transformer who holds capacity, sequencing and dependencies across competing priorities: who does what in which order, what waits, what gets escalated, and — the part no tool has ever captured — the relationships that make hand-offs actually happen.
Most of an Organizer’s work is invisible when it succeeds. The status that arrives before it is asked for, the conflict resolved in a corridor before it reaches a meeting, the dependency spotted two weeks before it bites. That invisibility is precisely why this role and agents are such a natural fit — and precisely where the danger sits.
What agents change for this role
The Organizer has the widest near-term scope for bounded delegation of all six roles. Routing, status-gathering, chasing, preparing, reminding: this is high-volume, repeating, rule-describable work, and agents do it willingly at any hour. Maren can delegate more of her week than Frauke ever could.
The failure mode is the mirror image of the gift: coordination becomes invisible, so nobody notices it has stopped working. When the chasing happens silently and the status assembles itself, the organisation slowly forgets that coordination is being done at all — until the week an agent quietly fails, or routes around a conflict no machine can see, and the project discovers that the glue was load-bearing. A human Organizer who stops working is noticed by lunchtime. An agent that stops chasing is noticed at the deadline.
In my work, this pattern has a name, and I learned it in my very first job out of university, when my boss explained watermelon reporting to me: green on the outside, red on the inside. Every internal team knows the truth — deadlines slipping, budget gone, bugs, technical debt — but towards leadership everything arrives green, because the right graphics and carefully optimised numbers can cover almost any red. The practice is as old as slides and survives every technology. It is also, in my view, both genuinely harmful and completely unnecessary: for a project or product lead, knowing what is going on before it blows up is not a nice-to-have. It is the job. The ones I have seen thrive — in product teams and delivery projects alike — were always the ones who leveraged the tools to know: the chat, the ticket system, the repository, and now agents, which accelerate that knowing beyond anything before. That is exactly what the status-evidence agent below exists for. When status assembles from the systems where work actually happens, and everything self-reported is labelled as self-reported, the watermelon has nowhere left to grow.
From one assistant to an Organizer’s suite
Maren’s meeting-preparation assistant is shape one. As the intake project spins up, the honest next step is separated purposes around her role:
- An intake and routing agent — watches the places where asks actually arrive: the project’s shared inbox, the chat channel where operations raises issues, the ticket system where IT logs change requests. Everything lands in one queue and is routed to the named owner, against a route Maren wrote down — so the integrator’s email no longer lives a separate life from the ticket it should have become.
- A status-evidence agent — assembles status from the systems where work actually happens: tickets closed in the project tool, changes merged in the integrator’s repository, test results filed by QA — and labels anything that exists only on a slide or in an email as self-reported.
- A dependency watcher — reads the plan and the links between tickets, knows which items block which, and flags when a predecessor slips: the integrator’s milestone moves, and the QA items depending on it light up the same day. It flags; it does not re-plan.
- A meeting-preparation agent — her original assistant, now fed by the other three: the agenda from last week’s minutes, what changed from the status evidence, the conflicts from the watcher.
- An escalation assistant — drafts the escalation the moment a trigger condition is met — a milestone slips past its buffer, a red sits unacknowledged for two days — with the case, the evidence and the options. A person sends it.
What is deliberately not on this list is a sequencing agent — something that re-orders the plan, re-prioritises between departments, or decides which conflict matters most. Sequencing is where capacity, politics and strategy meet, and the moment it is delegated, the Organizer has handed over the one judgement the role exists for. Priority conflict, relationship repair, the choice of which exception to fight for: these stay with Maren, and they are most of why she is good at her job.
Cost note: five agents means five briefs to keep true, five things to re-check when a route or an owner changes, and a coordination layer that itself needs coordinating — the suite pays for itself only because this work repeats every week.
Some coordination genuinely becomes a bounded lane — the intake project itself will produce one, with routine supplier orders classified and booked automatically and a person at the exception. That step raises its own demands: an explicit route, named owners on both sides of every hand-off, stop rules a working-level person can use immediately. The previous briefing sets out what a lane costs to run. The suite comes first, and most Organizers will not need to go further.
When access is the blocker
Two walls stand between this suite and reality in many firms. The first is policy: IT has a mandate not to connect AI tools to internal systems — often for good reasons, written after somebody’s shadow experiment touched data it should not have. The second is capability: neither Maren nor IT has ever wired an agent to a ticket system, and nobody wants to learn on a live project.
Neither wall means no start, because an agent does not need integrations to be useful — it needs material. Maren can begin with what she is already allowed to handle herself: a mailbox the intake agent reads because she forwards into it, a weekly export of the ticket list, the minutes she pastes in. The human is the integration. It is lower fidelity and it costs her minutes a day, and it is enough to prove within a month whether the suite deserves real connections. PathPatron runs on exactly this pattern — disconnected agents, a person handing files between them — and it publishes weekly.
The policy conversation changes shape too. IT’s mandate is almost always against uncontrolled tools, not against governed ones. So Maren does not arrive with a tool request; she arrives with the maps: what it would read, what is excluded, under whose identity it runs, who can stop it — and she asks what is already approved before asking for an exception. A request shaped like a boundary gets a different answer than a request shaped like enthusiasm. And the capability gap resolves in the order this whole series assumes: she explores on manual feeds, and hardening the connections is exactly the moment engineering is for.
The Organizer’s data and evidence map
Coordination agents read more of the organisation than any other kind, and that reach is where an Organizer’s evidence design earns its keep.
- The status source of truth. When the project tool and the team disagree, which is true? Declared in advance. The status-evidence agent labels self-reported status as self-reported, and conflicts surface — a green in the tool next to a worried message from the team lead is not noise, it is the finding.
- Reach and its limits. An agent that chases status can read calendars, message threads and ticket queues — which means it can, in principle, see how individuals work. The exclusion is drawn here, in advance: coordination agents report on work items, never on people. No response-time league tables, no inference about who is slow, no flagging of individuals — enforced by what the agent’s identity can reach and what its briefs forbid, and visible in its outputs, which name items and owners, not behaviour.
- Why this line is not optional in Germany. Maren works at a German firm, and that turns this exclusion from good practice into legal duty. Systems capable of monitoring behaviour or performance bring the works council to the table — co-determination is not a courtesy, it is law, and an Organizer who designs the exclusion first arrives at that conversation with the answer instead of the problem.
Across Europe: employee-representation regimes differ widely. Germany’s co-determination gives the works council a formal say when a system could monitor behaviour or performance; France’s CSE and the Dutch works-council model carry consultation duties of their own, while the UK has no directly comparable mechanism. The design principle travels even where the legal duty does not: involve the people whose work becomes visible, early. None of this is legal advice — the authority question, including what agents may do and under whose oversight, has its own briefing later in this Path.
How the first one actually gets built
Maren’s suite did not arrive as a project. The first agent took an afternoon, and the sequence is repeatable by anyone with the tools their firm already licenses.
- Pick the task you do weekly and resent. Hers was assembling the steering agenda. Not the most valuable task — the most repetitive one, because you will notice immediately whether the output is any good.
- Do it once, out loud, in writing. What she opens, in what order, what she copies from where, what she checks before sending. That page is the brief. Most of the work of building an agent is writing down work you have never had to describe.
- Give it the material by hand. Last week’s minutes, the current ticket export, the agenda template. No integrations, no permissions, no IT ticket.
- Compare against your own version for three cycles. Run it in parallel and fix the brief where the output disappoints — this is the loop, and three rounds is usually enough to know whether the thing earns its place.
- Then, and only then, ask for connections. Now the request has a track record, a written brief and a named boundary — the shape IT can say yes to.
The second and third agents are faster, because the hard part was never the tool. It was the description.
The questions that travel
Everything above compresses into four questions. Maren asks them of every coordination agent before it runs; they work just as well on a vendor’s “AI project manager” demo.
- Which hand-off does this touch, and who is named on each side of it? A good answer sounds like: “supplier order arrives by email, the intake agent files it as a ticket, Ayla in operations owns it from there.” A bad one names a department.
- What does it do when the route is unclear — queue, guess, or ask? Ask for the actual behaviour, not the intention: “unroutable items go to a review queue Maren clears each morning” is an answer; “it uses AI to determine the right owner” is not.
- Who notices if it stops working, and how soon? There has to be a person and a signal — “the daily digest arrives at 8am; if it is missing, Maren knows by 8.15.” If the honest answer is “the deadline”, the coordination has already gone invisible.
- What does it report about people? The right answer is nothing. If a report contains a person’s response times, reply rates or activity, it has crossed the line — items and owners, never behaviour.
In my work, the pattern when such a person leaves is depressingly consistent. First, a bit of panic. Then a high performer with ambitions is asked to fend it all off — on top of their own job, until a replacement is hired for one role or the other. And the history is lost anyway. I have seen it lost even behind the most diligent Organizers, people with epics and tickets properly described, acceptance criteria and all — because oversight does not transfer by existing, it transfers by being read, and nobody ever takes the time to read it manually. This is one thing agents genuinely change. Fed with the predecessor’s ticket history, the folders on their desktop, the scattered notes and half-finished documents, an agent makes the record usable instead of merely present: the successor can ask it questions on day one. Coordination knowledge used to die with handovers because reading was too expensive. It is not anymore.
There is an upside worth naming, because it is the part nobody plans for. If the person who left was good — or if the good one left three years ago — their record is not just history, it is a worked example of what strong coordination looks like. An agent fed with it can show a successor the pattern: how this team wrote up a dependency, what a well-formed escalation contained, which detail turned out to matter later. That is a real head start for someone with less experience, and it is the overloaded stand-in’s chance too: the person handed a second job at short notice can organise it with the same agents and come out looking, rightly, like they mastered it.
If the Organizer is on your team
Half the readers of this briefing coordinate. The other half lead a coordinator — a programme lead or project manager whose chasing and status work is increasingly done by agents. Your job is not to read the raw status yourself; it is to check that the coordination is still owned. Three checks do most of the work.
- Ask what the agents chase and what she decides. Chasing, assembling and preparing delegate well. If sequencing decisions or priority calls are arriving “from the system”, the role has been hollowed, not helped.
- Ask who notices when an agent fails silently. There must be a named answer — a person or a check, not “we would see it eventually”. The monitor needs a monitor.
- Watch the conflict signal. Healthy coordination surfaces disagreements; it does not dissolve them. When dependencies are “managed” but conflicts stop reaching you, the political work has been dropped, not delegated — and that work was the role.
One more duty sits with you, not with her. If Maren is hand-feeding agents from exports because no approved tools exist outside IT, no governed way to connect them, and no visible permission to work this way, that is not her workaround to admire — it is your access design to fix. Three decisions are yours:
- Tool provision — which agent-building tools are available beyond the engineers.
- Governance rails — which connections to internal systems are pre-approved, and under what rules, so a request shaped like a boundary has somewhere to land.
- Adoption comfort — how comfortable your non-technical departments are allowed to become with tech-driven daily work.
The manual-feed start is a proving ground, not a permanent condition — whether it stays permanent is a leadership choice.
Leading people who work with agents — the checks, the warning signs, the review cadence, for all six roles — is covered in its own briefing later in this Path.
The sequence has one name on it
Launch week arrives with a red flag: the integrator slips a milestone, and the dependency watcher flags that testing now collides with the ERP’s month-end freeze. The escalation assistant drafts the case within minutes — the slip, the collision, three options with their consequences. What happens next is nothing an agent does. Maren calls the quality lead before she calls the meeting, trades one test cycle against a staged go-live, and re-sequences the plan with her name on the change.
The chasing was agent-made, and everyone knew it. The sequence has one name on it, and everyone knows whose. That is the Organizer’s version of the series’ claim: agents can carry the legwork of coordination at a scale no person could — and the moment they carry the judgement, there is no Organizer left, only a very busy system that nobody is watching.
Related reading
- Delegate, Automate, or Keep Human: The Agent Handoff Map — assigning authority level by level across the Organizer’s suite.
- Map the Work Before You Automate It — the method behind every route this briefing assumes was written down.