BUILD FILE / COBALT · AUTONOMOUS GTM · 2026

We built a GTM engine that rejects its own work.

Six agents that research an institution the way a good analyst would, and one auditor with veto power over the other five. Built for Cobalt in fourteen weeks, then handed over whole.

6 agentsevery claim sourcedfully transferred

fig. 01 / one account through the pipelinerunning

idle · waiting for a trigger

the brief

Their buyer deletes anything that smells like sales.

Cobalt sells architecture intelligence into large financial institutions. The buyers are enterprise architects, CIOs and heads of AI. Deeply technical, no patience for marketing, and one templated email burns the sender permanently.

So the job was never “send more email.” It was to research an institution properly, find a real and recent reason to write to each specific person, sound like a founder rather than a tool, and refuse to send anything it could not prove.

prospector315 LINES

Watches for the moment an institution becomes worth approaching. It is allowed to find nothing, and often does. Silence beats a manufactured reason.

researcher462 LINES

Takes the institution apart, then each person. Dossier, org map, public record, a verified route in. Every fact labelled verified or hypothesis, nothing in between.

strategist518 LINES

Decides who is worth approaching, and matches each person to the angle their own record points at. Also decides to walk away.

writer728 LINES

Writes every message. The biggest brain in the system, and most of it is rules about what not to do. Sixty to eighty words, hard cap.

auditor270 LINES

Reviews all four and can veto any of it. Re-fetches sources live. Never rewrites: it blocks, explains, and the writer fixes it.

fixer91 LINES

A parallel lane for surgical jobs, so a one line fix never waits behind an hour long research run.

the research

It writes to what a person actually said. Not to a job title.

Finding a company is easy. Finding a real, recent, provable reason to write to one specific person inside it is the whole game. That is what the researcher does, for every prospect, before a single word gets written.

Exactly which signals it hunts, and where, is the client’s edge now, so this page keeps that to itself. What we can show is the standard every fact had to meet before anyone was allowed to use it.

standard

Sourced, or silent

Every hook names where the thing was said and carries the live link. Inventing a reason to write is never an option. Finding nothing is.

standard

Recent means recent

Every fact gets its date checked and its age calculated. Stale material is not a trigger, and older context has to be framed as old.

standard

Verified or hypothesis

Every claim in the dossier carries one of two labels, and a hypothesis is never allowed to reach a prospect dressed as a fact.

the audit

The agent that writes is not the agent that approves.

Deep research creates a new risk. The more a system finds, the more confident and the more wrong it can be. So nothing reaches a person until a separate agent has tried to take it apart.

The auditor re-fetches every source, live, and checks that the specific claim actually appears where the writer says it does. A source that leads somewhere plausible but unsupported is treated as fabrication, because to the person receiving it there is no difference.

When it blocks, the writer fixes it and resubmits. And if the two keep disagreeing, the system stops arguing with itself and asks a human.

verdicts

PASS

Sources re-fetched and confirmed, copy within limits, no fingerprints. Moves to the human queue.

PASS WITH FLAGS

Sendable, but the auditor wrote down what it was uneasy about, and the flag travels with the message.

BLOCK

Stops. Specific written reason, back to the writer, full history kept.

things it actually caught

×A source attributed to the wrong article

×A name collision that had pulled research on an entirely different person

×Messages still assigned to a sender who had left the company

×A claim whose linked page was real, but did not say the thing

Every one of those would have gone out.

what it added up to

Real counts, from the production database.

The engine came with a full control room, because autonomy without visibility is just risk. Every claim in the interface carried a clickable source, nothing sent without a person approving it, and every correction someone made was diffed, logged, and read back by the agent that made the mistake.

0institutions mapped and tiered

0researched end to end, dossier verified

0decision makers identified

0buying triggers found, each one sourced

0messages written and reviewed

0corrections absorbed by the agents

0lines of control room behind it

0sent without a human approving it

also in the box

That was the spine. All of this shipped around it.

agent chat

Ask any agent why it did something, straight from the dashboard. Pin an instruction and it becomes a permanent rule on every future run.

one-click fixes

Find an email, rewrite one message, add one step. Buttons in the dashboard, handled by the surgical lane in minutes.

independent checks

After every stage, the system verifies the work is really there before anything moves downstream. A failure halts loudly.

meeting briefings

One click before a meeting produces a researched brief on the people who will be in the room.

calls filed to CRM

Every recorded call is transcribed, summarised and filed against the right company in their CRM. Hourly, hands-free.

reply-aware sending

Email and social sequences wired into the client’s own tools. The moment a prospect replies, their sequence stops.

per-account pages

Personalised one-page sites for the highest-value accounts, with visit tracking wired back into the system.

live agent status

The dashboard shows what every agent is doing right now, and anyone on the team can open an agent’s brain and read its actual instructions.

cost ledger

Every pipeline run records what it cost, agent by agent. The owner always knows what a researched account actually costs.

the ending most agencies skip

Then we gave it away.

An engine you cannot own is a subscription with extra steps. So the last phase was handing over every part of it, to their accounts, under their control, with us holding nothing back.

ServerThe machine the agents run on

transferred

DatabaseEvery institution, person, sequence and dossier

transferred

RepositoryAll six brains, the watcher, the control room

transferred

Deployment pipelinePush to their repo, deploys to their hosting

transferred

Model accessMoved to their own seat, so every run bills them

transferred

The person who runs itCoached through the build from day one, owns it now

in house

The engine kept running through every one of those transfers. Zero downtime, verified end to end afterwards.

The full build file

Unlock the whole system breakdown

The architecture, the agents, the numbers, and what the client owns now. Enter your work email and it opens instantly.

Free. No spam. One email from Amit, maybe.

what this actually is

A campaign ends. This does not.

Cobalt did not buy a list, or a sequence, or a month of someone else’s attention. They own a research and outreach machine that runs on their own infrastructure, holds everything it has ever learned about their market, and gets sharper every time a person corrects it.

This is what Altitude builds. Not a service you rent. A system you keep.

Build one for us →