SD Agentic
Founders' plan · Nick & Jonas
That code didn't work. Please try again.
Founders only · Confidential
Founders' plan · Nick & Jonas
That code didn't work. Please try again.
Founders only · Confidential
A plan to put agents to work across Saigon Digital: the marketing team we don't have, client delivery at scale, and an operation that runs on SD Friday instead of Nick's laptop. Nick and Jonas decide and specify it. Tri's team builds it. People stay in charge of everything that leaves the building.
Our best automation is sales tooling that lives on one Mac. Delivery and SD's own marketing are almost entirely manual.
Every agent reads and writes through SD Friday. Every output is a record with an owner and an approval.
Growth, Delivery and Ops. Six agents already exist in some form, ten are new.
Make 8 decisions and write 5 agent charters. Then Tri builds in phases, first results in about 6 weeks.
Pulled live from SD Friday on 17 September. The pattern is the same everywhere: good systems, too few hands, and records that drift out of date because updating them is nobody's job.
SD's monthly social content task is on hold. We've produced 90 proposals and 142 AI visibility audits, and turned none of them into case studies or content.
Prospecting, outreach, audits, proposals and reply handling all run through Nick plus Claude on one laptop. It works. It also stops when Nick does.
SEO blogs and monthly reports queue up in "Ready for QAQC" across the SEO team. The automated report pipeline covers one client.
Sources: SD Friday getSalesFunnel, getFinancialMetrics, getOverdueInvoices and getTeamWorkload, 17 Sep 2026. Win rate = Closed Won ÷ (Won + Lost) by count. The shared "Developers" account holds another 281 tasks.
A sweep of our own repos. Sales automation is genuinely strong. The problem is where it runs and how little of it touches delivery.
Projects, tasks with review and QA roles, leads, contracts, invoices, SEO KPIs. An MCP server with read and write tools, a Claude chat, AI briefs for platform requests, auto-drafted contracts, weekly health reports.
Audit tool: prospector and lead-gen agents, enrichment, email verification, ReachKit campaigns, daily cold drafts, reply detection, AI visibility audits. Skills for warm replies, call scripts and proposals.
Proposal Hub with grow. and build. decks and open tracking, the deck engine, 90 web proposals, the Go-Live Checker for launches.
Monthly report pipeline: GSC, GA4, Ahrefs and CrUX pull, 13 QA checks, narrative, web page and PDF. Set up for VALO only. The SEO Command Center is a separate, script-built view.
No brief to draft to QA to publish pipeline for clients, even though SD Friday already models peer review, QA and publisher roles on tasks.
About 30 marketing skills are installed and unused. No blog, LinkedIn or case study pipeline. We don't track SD's own AI visibility, which is the thing we sell.
Automation is split across Nick's Mac, Railway, Vercel crons, Netlify and local scripts, with about 40 one-off campaign fix scripts and no shared queue or monitoring. Two leads pipelines exist (the audit tool and SD Friday). We also found credentials stored in plain files in a few repos. Phase 0 fixes the foundations before we add more agents.
The same platform we're now pitching to clients like 7 Bridges, built for ourselves first. If it works for SD, it becomes a product.
Events, schedules and requests
Single source of truth
One runner, one queue, one log
Nothing external without a yes
Logged back to SD Friday
Each one has a job, reads and writes named SD Friday records, and stops at a named person. Filter by squad.
Works the 35 proposals sent and 115 leads sitting at Lead In. Drafts chasers, flags stalled deals, builds reactivation lists from 547 closed-lost leads.
Two SEO and AI-search articles a week for saigon.digital, plus LinkedIn posts in Nick's and Jonas's voices, built from real work and GSC gaps.
Turns closed-won projects, monthly reports and audits into case studies and anonymised proof points for proposals and the site.
The audit tool's prospecting, enrichment, verification and ReachKit campaigns, moved onto the shared runner and writing leads straight into SD Friday.
A reply comes in, the agent builds the audit, drafts the response, files the lead at the right stage and posts a card to the sales chat.
From a lead record and call notes to a branded proposal, with the 60% margin check run before anything is sent.
When a contract is signed, turns scope into a client brief, a task plan with reviewers and dates, and access checklists.
Brief, draft in English or Vietnamese, fact and brand checks, structured data, then into the task's peer review stage. Writers become editors.
The existing report pipeline rolled out to every SEO retainer, with the narrative drafted automatically and flagged where the data looks wrong.
Watches client sites for ranking drops, broken pages, blocked AI crawlers and slow Core Web Vitals. Opens an issue before the client notices.
Drafts a short weekly update per client from what actually moved in SD Friday that week, in the client's language.
Claude Code set up for Tri's team: component scaffolds from design, content migrations, redirect maps, visual QA before launch.
One page every Monday: revenue, pipeline movement, cash, capacity, at-risk projects and the three things that need a founder this week.
Finds stale tasks, duplicate tasks and projects "past deadline" that are really ongoing retainers. Proposes closures in batches.
Reconciles the 150+ "overdue" invoices against payments, then chases only what is genuinely owed, with a polite reminder ladder.
Matches pipeline likely to close against team hours, warns before we overcommit, and tracks real margin per project against the 60% target.
Every one of these comes from something that actually went wrong in our own automation this year. They go into every agent's charter.
No agent sends email, publishes or messages a client. It drafts, and a named person approves.
A Gmail "reply" tool sent a prospect follow-up that was only meant to be drafted.Each agent runs on exactly one server, with a single switch to turn it on or off.
A second copy of the pipeline on Railway re-audited leads that had already been emailed.Suppression and "is this our own client" checks exclude on any doubt.
Tightening a matching rule let an existing contact through and caused a live opt-out.Every new alert runs silently for a week, and we check its hits before anyone gets pinged.
The hot-reply alert fired eight false alarms on its first run.Each run records inputs, outputs, cost, approver and outcome against a lead, project or task.
Today there is no single place to see what our automation did yesterday.No keys in repos or files. Each agent gets the narrowest access it needs.
The repo sweep found credentials in plain files in several projects.Phases start Monday 21 September 2026. Each phase ends with something working that we use every week, not a demo.
Done when: Tri has five signed charters, access, and a clean board.
Done when: Nick and Jonas get the Monday Brief and approve chasers from the inbox.
Done when: saigon.digital publishes weekly without a marketing hire.
Done when: the "Ready for QAQC" queue is reviews, not writing.
Done when: it's a line item in proposals, not a slide.
One page per agent, written by Nick or Jonas in Phase 0. Filled in here for Pipeline Nurse.
| Goal | No proposal goes 7 days without a follow-up. Revive 5 closed-lost leads a month. |
| Trigger | Daily 7am, plus any lead moving to Proposal Sent |
| Reads | sales_leads (stage, last activity, value), campaigns, the Gmail thread, the proposal link and its open tracking |
| Does | Scores each open deal, drafts a chaser in the thread with a reason to reply, suggests the next step |
| Writes | Gmail draft, a lead note with its reasoning, an updated next step |
| Human gate | Nick approves or edits in the approvals inbox. Nothing sends automatically. |
| Never | Contact suppressed domains, current clients or leads Nick is handling personally |
| Success metric | Days since last touch on open proposals, and win rate on Proposal Sent |
| Cost cap | Monthly AI spend limit set by the founders, with an alert at 80% |
| Owner | Nick (business), Tri (technical) |
Nobody's job disappears in this plan. The work moves from producing to directing, reviewing and approving.
Stops being the machine. 20 to 30 minutes a day in the approvals inbox for outreach, content and proposals.
Owns data quality in SD Friday, approves receivables and capacity calls, and reads the Monday Brief.
Builds and runs the spine and the agents. Needs protected time, which starts with clearing a 252-task queue.
Approves kick-off plans and margin checks. Reviews quality across the delivery squad.
Sends client updates and reports the agents draft. Checks consent on case studies.
Review and improve agent drafts, own the brief and the quality bar, and spend saved hours on strategy.
Illustrative scenarios using today's SD Friday numbers. They are not forecasts. The point is how little has to improve to pay for this many times over.
| Lever | Today | If the agents move it | Worth |
|---|---|---|---|
| Proposals sent that close | 35 open, USD 1.04M | 5 more percentage points close | ≈ USD 52k |
| Leads parked at Lead In | 115 leads, USD 1.97M | 3% worked through to a win | ≈ USD 59k |
| Closed-lost reactivation | 547 leads, USD 5.25M | 1% come back | ≈ USD 52k |
| Real receivables recovered | Unknown until cleaned | Chase only what's genuinely owed | Cash, not revenue |
| SEO team hours | Blogs and reports written by hand | Writers edit drafts instead | More clients per head |
| SD's own marketing | On hold | Weekly content without a hire | Inbound we don't have today |
The real investment. Roughly half of Tri's week for about 12 weeks, plus dev support. That only works if the stale queue is cleared and something is deprioritised.
A monthly cap the founders set, tracked per agent in the run log. Start small with the first three agents, and raise it only when an agent proves its value.
One small always-on server or the existing Railway/Vercel setup, plus a secrets vault. Small next to the time cost.
252 active tasks. Without protected time, this becomes a side project that never ships. Mitigation: Board Hygiene first, and a founder-agreed cap on client work.
Agents acting on stale invoices and tasks will embarrass us. Mitigation: Phase 0 cleanup, and Jonas owns data quality.
Our reputation is the product. Mitigation: the draft-first rule, named approvers, and every alert tested silently first.
If Nick gets 60 approvals a day, we've moved the problem. Mitigation: batch approvals, and loosen gates only on agents with a clean record.
Usage grows quietly. Mitigation: cost logged per run, caps per agent, and a monthly review in the Monday Brief.
Editors who don't trust drafts rewrite them from scratch. Mitigation: start with agents that remove chores, and involve Simon and Kes in the quality bar.
Each one has a recommendation. Agree or change it, and Phase 0 is half done.
Without a number, client work always wins.
Recommend: 50% for 12 weeks, protected.Nick's Mac, Railway, Vercel, or a dedicated box.
Recommend: one always-on server, never a laptop.Chores that pay back fast, before anything client-facing.
Recommend: Monday Brief, Board Hygiene, Pipeline Nurse, SD Content Engine, Monthly Reporter.A number that makes experiments safe.
Recommend: start with a small cap for Phase 1, review it monthly.Agents need someone who owns the voice and the calendar.
Recommend: agents plus a part-time editor role, decided after Phase 2.What client data can go into which models, and what clients are told.
Recommend: a one-page policy before any delivery agent goes live.Named gates per squad, with a backup when someone is away.
Recommend: the five approvers on this page, each with a named backup.Sell it to clients from Q2 2027, or keep it internal.
Recommend: yes, once our own numbers prove it.Ninety minutes this week: go through the eight decisions, pick the first five agents, and assign who writes each charter. By 2 October, Tri has everything needed to start.