Warm-path precision

The warmth graph, on the world

Live proposal · synthetic data
Warm-path precision

Who do we know
at Northwind Water?

Every person we deal with, placed where they are on Earth. The glow of each arc is the warmth of the tie — so the answer to "who sends the email" is visible, not guessed.

Organisation

Warm path
ColdWarm
Warm tie Latent (LinkedIn) Cold
Drag to spin · click a contact

The demo Karl built is a catalogue of agents. What we're proposing is the layer underneath it: graph engineering — many agent loops wired by edges, checked by verifiers, and pinned to anchors the machine may never rewrite. That last rule is what keeps a fleet of 60 agents honest instead of agreeing itself into confident nonsense at scale.

61Agents
10Departments
Start at L0 · observe
3Live today

The fleet — every agent is a governed loop

Karl's 61 agents, kept exactly — plus the 10 AOS warmth agents (first, in orange) that build the graph the globe runs on. Every one is a loop: pinned to an anchor, gated on an eval, promoted up the Trust Ladder only on evidence. Click any agent.
L0observe L1draft L2act·reversible live partial planned

The graph — one fleet, many outcomes

Karl needs several outcomes. Each is a different graph wired from the same pool of agents — switch between them. Agents reused across graphs are marked . Warmth agents are orange, Karl's blue, gates green.

Every agent learns — and we can prove it per agent

The loop closes: every reply and meeting scores the agent that produced it, so each one gets better on its own. Because it’s measured per agent, we can see exactly which loop improved — and the graph’s yield compounds. Reply yield 2.4% → 6.1% over 8 weeks.

Anchors — the parts the optimiser may never rewrite

Facts, evals and audit records with a human owner. This is the defensible core of the design.
🔒Ledger — append + read only 🔒Warmth holdout — sealed, salted 🔒Corpus — cited knowledge 🔒Suppression list 🔒Audit
The invariant: the optimiser may rewrite any loop's prompt, skills, thresholds or edges — but it can never write an anchor. Anchors change only when a named human rotates them. That single rule is the Goodhart defence, and the most defensible thing we own.

Authority is earned — the Trust Ladder

Every agent starts at the bottom. Nothing sends on its own until it has the track record.
L0 · OBSERVE

Read & report

Can surface, summarise and rank. Cannot produce anything that leaves the system. Every new agent starts here.

L1 · DRAFT

Produce, never send

Writes emails and assets, staged for a human. Promotion needs a passing eval suite.

L2 · ACT (reversible)

Change what can be undone

Update CRM, file a doc, open a PR. Needs 50+ clean runs at L1.

L3 · ACT (irreversible)

Send · deploy · spend

A named human authorises every time. No promotion path to unattended.

What the graph decides: cold vs warm

People send the warm ones

The daily 15

A strong tie on the globe → the draft lands in your own Outlook, and you press Send. ~15 a day keeps sender reputation pristine and every send a genuine human action.

  • From your own name and mailbox
  • Personalised from a real trigger + the relationship
  • Spam-Act clean — a person sending their own email
Agents send the cold ones

The autonomous fleet

No tie on the globe → routed to agent-owned marketing addresses that auto-send inside a human-approved envelope — never impersonating a person.

  • Company marketing identity, not a real person
  • Caps, cadence & circuit-breakers on every sender
  • The earned L3 exception, owned by Phil + Nolene

The operational picture — accountable to 50 meetings a week. Warmth precision and Karl’s fleet, measured on one board.

This week · live
Meetings booked
34 / 50
run-rate 7.1/day · need 8.0
Meaningful interactions
168 / 200
▲ +26 vs last week
Positive responses
96
57% of interactions
Pipeline influenced
$2.4m
9 new opportunities
Required run-rate
8 / day
to reach 50 by Friday
Forecast (EOW)
47 · 82%
trending toward target

Production funnel

actual vs required · this week

Live operations

streaming

Sender health

12 of ~50 mailboxes

Workflows in production

the graphs, running