Client LinkedIn Ads · Dust build · Workflow & route decision
The full workflow we can build, marked for where an agent runs it, where it drafts and you approve, and where a human simply has to act. Plus the two ways to build it — in the client's Dust now, or yours first.
Context: Running LinkedIn ads for a client who already uses Dust. Goal: let agents do as much of the create-and-manage cycle as possible — thought leader ads, organic posting from 3 profiles, message ads to prospect lists, company-page ads, and reporting. Human touchpoints are welcome where needed. Decision pending: build in their workspace, or build in mine and copy across.
At a glance:
One pass through the engine. The brain (Dust) reasons, drafts and reports; LinkedIn's rules decide where a person has to step in. The two stages LinkedIn won't let software touch reliably — posting from personal profiles and building message ads — stay human by design, not by limitation of the build.
Legend: Agentic = agent runs end to end. Draft to approve = agent prepares, you sign off. Human = a person must do this.
The flow, in order:
Running continuously alongside: the agent watches spend, flags drift, and proposes budget moves — you approve any spend change.
What each stage does, who runs it, and what LinkedIn actually permits via its Marketing API. The constraint noted is the honest limit — most are platform policy, not build effort.
Stage 1 — Research. Mode: Agentic. Agent reads the client's CRM, Drive and past campaign data; builds ICP and a brief. Constraint: none — runs on connected data sources.
Stage 2 — Creative. Mode: Draft to approve. Agent drafts TLA posts for the 3 leaders, company-page ad copy, and message sequences. Constraint: brand voice and legal sign-off need a human eye.
Stage 3 — Organic posting. Mode: Human. The 3 thought leaders publish from their own profiles; the agent supplies final text and a queue or reminder. Constraint: personal-profile posting via API is effectively closed to third parties, and session-based auto-posters risk the account.
Stage 4 — Audience. Mode: Agentic. Agent assembles target lists from CRM segments, formats them, and uploads as matched audiences. Constraint: list and contact upload is API-supported; one human GDPR check before push.
Stage 5 — Campaign build. Mode: Draft to approve. Agent constructs company-page Sponsored Content and TLA sponsorship: structure, targeting, budget, creative. Constraint: fully API-supported but needs Marketing Developer Platform approval (about a 2–4 week review). You launch.
Stage 6 — Message ads. Mode: Human. Agent drafts the conversation flow; a person builds and launches it in Campaign Manager. Constraint: the Sponsored Messaging API is private (extra approval) and EU recipients must opt in first. Treat as manual.
Stage 7 — Monitor and report. Mode: Agentic. Agent pulls analytics daily, flags drift, proposes budget moves, and writes the report to Slack or Notion. Constraint: the Ad Analytics API is built for this; spend changes route through you.
A small orchestrated team, not one giant agent. One coordinator routes work to four specialists. Mirrors the Sam-and-sub-agents pattern, scoped to this engine. Each specialist owns one job and a defined set of tools.
Coordinator (orchestrator). Owns the cycle. Takes "run this week's push", routes to specialists, collects drafts, presents them to you for approval, and never launches anything itself. Tools: none autonomous; human-gated.
Researcher (agentic). Reads CRM, Drive, past performance. Produces the ICP, segment definitions and a creative brief the copy agent works from. Tools: CRM, Drive, Analytics.
Copywriter (draft). Drafts all creative to brand voice: TLA posts per leader, page-ad variants, message sequences, organic posts. Output is always a draft for review. Tools: voice guide in, brief in, draft out.
Audience builder (agentic). Turns CRM segments into LinkedIn-ready matched audiences, formats and stages the upload. Pauses for your GDPR confirm before pushing. Tools: CRM, LinkedIn API.
Campaign ops (draft). Builds page-ad and TLA-sponsor campaign objects via the Marketing API: structure, targeting, budget, creative assignment. Stops at "ready to launch". Tools: LinkedIn API; launch is human.
Analyst (agentic). Daily analytics pull, anomaly flags, budget-shift proposals, and the written weekly report posted to Slack, Notion or Drive. Tools: Analytics API, Slack/Notion.
Four points keep a person in the loop. Two are platform rules we can't engineer around; two are deliberate control gates we'd keep even if we could remove them.
The brain is just instructions and structure, so most of it can be built and tested today on stand-in data. Only the live plumbing waits for their workspace.
Can build now:
Waits for their workspace:
Build-for-portability rule: keep every account-specific value — ad-account ID, list names, page IDs, connector references — as a parameter at the top of each agent, never buried mid-prompt. Then moving to their workspace is "change these six values", not "re-read the agent". Keep a one-page list of every external connection each agent touches; that becomes the wiring runbook.
You asked for the case for each. Here it is straight — then my recommendation, which is neither pure option.
Request access, connect their tools, build live.
For:
Against:
Design and test on your Dust, recreate in theirs.
For:
Against:
Don't pick one. Build the portable brain — researcher, copywriter, analyst, orchestration, report format — on your Dust now, because none of it needs their access and all of it becomes reusable DiscoSam IP. In parallel, request access to their Dust for the plumbing-dependent stages (audience push, campaign ops, live connectors) and do the final wiring there. You design on yours, deploy on theirs. The dev never blocks on the client, and the handover is a wiring session, not a rebuild.
Four things to settle before we commit the build.
Who applies for the Marketing Developer Platform — you or them? The campaign-build and reporting stages need MDP approval (about a 2–4 week review against an established business and use case). Their entity may already have it, or we ride an approved partner wrapper to skip the queue. This gates Stages 4, 5 and 7. Decide with the client before build kick-off.
Whose workspace owns the engine long-term? If it lives on their Dust, they own and pay for it. If you want it as a repeatable product, the brain stays yours and only a deployed copy sits with them. Affects IP and how you price. Decide yourself this week.
Are the 3 thought leaders bought in to a posting cadence? Stage 3 only works if they'll publish on schedule. Auto-approval for sponsorship can be switched on once, but the posting itself is on them. No buy-in, no TLA engine. Decide with the client before Stage 3.
How much of the audience is EU-based? Determines whether message ads are viable at all. Heavy EU targeting means Stage 6 stays small and opt-in-gated; we lean budget into retargeting engaged TLA viewers instead. Decide with the client during planning.
What could break or bite, and how we hold it.
MDP approval delays or is refused. Severity: high. Mitigation: ride an approved partner API wrapper to skip the queue, or have the client apply under their established entity. Build the brain in parallel so we're not idle.
"Works on mine, breaks on theirs" at handover. Severity: medium. Mitigation: parameterise every account value; keep a connection runbook; budget a dedicated wiring and re-test session rather than promising a flip-the-switch migration.
Personal-profile auto-posting tooling risks the accounts. Severity: high. Mitigation: don't automate Stage 3 posting at all. Agent drafts and reminds; the leaders post manually. Three accounts triples the ban risk if ignored.
Message ads to EU prospects breach opt-in rules. Severity: medium. Mitigation: keep Stage 6 human-built and compliance-checked; default EU budget to retargeting, not cold messaging.
Thought leaders don't post on cadence. Severity: medium. Mitigation: confirm buy-in up front; agent runs reminders and a ready-to-paste queue to lower the friction to near zero.
Agent reports fabricated success (Sam-style failure). Severity: medium. Mitigation: the analyst must cite the analytics pull it's reporting from; no number without a source call. The human approval gate on any spend move catches the rest.
Client LinkedIn Ads · Dust workflow and route decision · 22 May 2026