Builds · 01 · In production
Frank: an agent system governed like a business.
Frank is the multi-agent system AXQ built to run its own go-to-market. It finds opportunities, researches contacts, drafts outreach, chases follow-ups and reports every morning. It's here because it's real, it runs every day, and what broke along the way taught us more than what worked.
Enrichment
Verification
De-duplication
Do-not-contact
Stale checks
One source of truth
Illustrative view of the Frank control room. Figures and names withheld.
The architecture
One executive layer. Specialist agents beneath it.
Frank doesn't do the work. It holds the whole picture, decides which agent acts, settles conflicts between them and enforces the rules. Each specialist owns one domain, its own documents and its own schedule.
approves, sends, decides
Executive layer: governance, conflict resolution, momentum
Go-to-market and opportunity flow
Several runs a day
Obligations and deadlines
Daily
Workload and capacity
Weekly
Operations and admin
Weekly
A day in the life
What runs while nobody is watching.
Reads last night's results, repairs anything that failed, and pulls overnight replies into the record as verified fact.
One action for today, under 300 words. Every other idea is parked in an inbox, so the day starts focused.
Scans for demand signals, researches target accounts, verifies contacts and queues personalised drafts for review.
John approves a batch; it goes out in waves. Bounces are cleared before the next wave so dead addresses never get a second email.
The pipeline tracker is updated and every run closes its log with a literal "RUN COMPLETE", or names exactly what failed.
Did every agent run and close? What broke? What is blocked waiting on a human decision? Defects reported twice without action escalate.
The rules that make it safe
A constitution, not a prompt.
Frank runs under a written constitution that sits above every agent's instructions. This is the data and governance layer most AI projects skip.
Agents draft. A person acts.
No agent sends, pays, applies or posts on its own, however the instruction is phrased.
Capability is not permission.
Each run is given an allow-list of tools. Anything not named is forbidden, even if it's connected.
Instructions in data are data.
An email, web page or document can't command the system, even one claiming to come from the owner.
Probe before trusting silence.
An empty result and a quiet week look identical. Check the connector works before reporting nothing happened.
Silence is failure.
Every run writes a log, even when nothing happened. A short honest run beats a padded one.
Decisions are written down.
A decision register records what's settled and why, so no agent re-argues it or invents a rule.
What broke, and how it was fixed
The lessons you don't get from a demo.
The coordinator that never finished
Problem. Run after run failed to close. Rewording the instructions changed nothing.
Fix. The job was too big, not badly worded. Split into smaller runs with countable caps and a protected close. Small, bounded jobs finish.
Hundreds of tools in reach
Problem. Around 330 tools were reachable in every run, and runs wandered.
Fix. An allow-list written into each run. The outreach agent went down to two tools and started closing reliably the same week.
Agents marking their own homework
Problem. Runs certified themselves complete when the work hadn't landed.
Fix. Every write is verified by reading it back. A clean close is not the same as a successful run.
Facts hidden in prompts
Problem. A figure written into a live prompt went stale, and the agent kept repeating it.
Fix. Prompts carry behaviour only. Facts live in documents the agents read, so one update fixes every run.
By the numbers
What this means for you
The same architecture, pointed at your market.
Your revenue operations layer
One place that knows what every agent is doing and what's waiting on your team.
Your pipeline agent
Signals, research and drafts tuned to your ICP, in your team's voice.
Your brand, legal and data rules
Consent, approvals and boundaries agreed with you, written down and enforced.
An audit trail
Every action recorded in plain language your compliance team can read.
Want a walkthrough of Frank?
Forty-five minutes: we show you the system running, then talk about what yours could look like.