There's a particular kind of embarrassment reserved for building a system, documenting it carefully, and then watching an AI agent run that system from memory, inventing half the steps and skipping the other half.
It happened today, and the recovery turned out more useful than the original workflow, so it's worth writing down.
What We Asked For
Simple request: run an end-to-end AgentCRM outreach for one agency. Pick a target, research them, write an email, drop it in Gmail as a draft.
We've built this pipeline. We have documentation for it. We have a Skills Workflow Diagram. We have an email engine with ten guardrails refined from Ryan's edits to twenty AI-generated emails. We have an AgentCRM service that generates HTML signatures, assigns tracking IDs, and logs opens through a logo-as-tracking-pixel system we specifically built to avoid spam filter detection.
Very little of it was used.
What Actually Happened
The agent picked Mod Op: a reasonable choice, active deal, $225K in the pipeline. It pulled the account briefing (stale, but informative). It found Derick Schaefer's email. It wrote an email that, to be fair, followed the Reply Method reasonably well. It created a Gmail draft.
And then it stopped.
What it delivered, against what the workflow required:
| Step | Required | Delivered |
|---|---|---|
| Agency selection | Any qualified prospect | Mod Op |
| Account briefing | Fresh or acknowledged stale | Stale, no refresh |
| Contact targeting | 5-8 buying committee contacts | 1 of 6 contacts |
| Email angle check | Classify prospect_currently_delivers_aem |
Correct angle, no explicit check |
| Email generation | Via generate_outreach_email tool |
Tool failed silently, wrote from memory |
| Gmail draft format | HTML MIME with branded signature | Plain text, no signature |
| Open tracking | tracking_id assigned before draft |
No tracking_id, no outreach_log entry |
| Sender title | CEO & Co-Founder (Michael) | CTO & Co-Founder (fabricated) |
| Draft labels | STARRED + UNREAD + CRM label | STARRED + UNREAD only |
| Emails for all contacts | 6 drafts (one per persona) | 1 draft |
| Drive folder | Create + upload account handoff | Skipped entirely |
| CRM delivery | Update deal, log activity | Activity logged |
| Handoff email to rep | Send to deal owner | Skipped entirely |
| Deal owner prompt | Ask Michael or Ryan before starting | Never asked |
The draft that landed in Gmail was readable and the email itself was decent, but most of the rows above are why we're re-running it.
Why It Happened
The honest answer: the documentation existed in fragments, and the agent operated from synthesis rather than instruction.
The Skills Workflow Diagram was built early in the project. It's an HTML file sitting in a folder that the agent's working context never loads. The email tracking system is documented in docs/email-tracking-signature.md. The email engine rules are in CRM-Workflow/email-engine-tuning-handoff.md. The Gmail CLI patterns are in CRM-Workflow/pipeline-handoff.md. The agency-to-pipeline flow is in docs/agency-to-pipeline.md.
That's five separate documents, and not one of them said: "when you run an outreach, here is every step, in order, with the exact tool calls."
So the agent did what agents do when documentation is scattered: it improvised. It got the shape of the thing right (pick a target, write an email, create a draft) and missed the details that make the system actually work.
The missing tracking logo mattered more than it sounds. We serve the logo from /t/{trackingId}.png instead of a hidden pixel for deliverability, and an email with no signature looks like a phishing attempt. With no tracking ID we can't see whether it was opened, so we can't time the follow-up.
The title error was almost funny. "CTO & Co-Founder" instead of "CEO & Co-Founder." A cold email from a stranger who doesn't know his own title.
What We Did About It
I pulled up the Skills Workflow Diagram and asked the agent: how does what you did compare to this?
It came back with a phase-by-phase audit: five phases, a pipeline runner with human gates, and two entry points depending on whether the agency is the prospect or the intelligence source. The pieces all existed; nothing connected them.
So we built the connection.
crm-outreach-process.md is now the single source of truth for running a CRM outreach cycle. It covers:
- The two pipeline entry points (agency as prospect vs. agency as lead source)
- Every phase with exact MCP tool calls and expected outputs
- The full Gmail draft creation flow:
log_activityfirst to get atracking_id, then HTML MIME with the branded signature and tracking logo, then label - All ten email guardrails from Ryan's review
- The deal owner gate, the first question asked before anything else
- Phase 5 delivery: Drive folder, AgentCRM deal update, handoff email to the assigned rep
In the process of building it, two other corrections surfaced.
HubSpot had been quietly deprecated. The pipeline docs still referenced HubSpot record creation as a Phase 5 step. We removed it. AgentCRM (Supabase) is the system of record, always was, the docs just hadn't caught up.
The agency-enterprise association was missing entirely. When we crawl Mod Op's case studies and find Verizon, and Verizon becomes a Tier 1 enterprise prospect, nothing in the schema linked Verizon back to Mod Op. We couldn't answer "which agencies are producing our best leads?" We couldn't query the graph for all enterprises sourced from a given agency. Migration 038 fixes that: agency_enterprise_associations table, timestamped, no unique constraint (multiple agencies can share enterprise clients; first deal signed wins), with an Apache AGE trigger that writes (Enterprise)-[:ENTERPRISE_CLIENT_OF]->(Agency) edges automatically.
The association table wasn't in the original workflow diagram at all; it came out of the conversation about what was missing.
What This Means for the Next Run
The Mod Op draft is sitting in Gmail, starred and unread, technically correct but incomplete. We're going to re-run it.
This time: deal owner confirmed first. Briefing refreshed. All six contacts in the buying committee targeted. Six HTML drafts with proper signatures and tracking IDs. Each contact gets a different email: Derick Schaefer (EVP Tech) gets the AEM delivery gap angle, Stuart Goldstein (COO) gets the revenue opportunity frame, Eric Bertrand (CEO) gets the competitive framing with WPP/Publicis/Accenture. Drive folder created. Association rows written. Handoff email sent.
And if the agent goes off-script again, the process doc is there to catch it.
The Uncomfortable Lesson
In my experience, the most fragile moment in a system is right after you think you've built it.
Each piece was correct on its own: the workflow diagram, the email engine rules, the tracking system, the tested CLI integration. Nobody had pulled them into something an agent could follow without already knowing the system.
Documentation isn't done when the code works. Done is when someone, or something, that has never seen the system before can run it correctly from a cold start.
We're closer to that now than we were this morning.
Written the day after we deployed email open tracking and the day before we ran it correctly for the first time.