How the agents work together — and with a human
A single morning in Georgetown, TX: six specialist agents hand work between each other overnight, then bring one decision-ready recommendation to a regional VP.
The agent team
Each agent owns a narrow job and defends its own numbers.
Watches permits, migration, real estate, construction and search demand across 1,284 markets.
Projects passings, penetration curves, mobile attach and 5-year revenue per neighborhood.
Tracks AT&T Fiber permits, Google Fiber builds, T-Mobile FWA and local ad surges; forecasts revenue at risk.
Costs the build, tests cost-per-passing thresholds and sequences construction against crew capacity.
Designs acquisition and defensive offers, models churn impact and margin dilution.
Arbitrates between agents, resolves conflicts, writes the recommendation and brings the human in.
The human in the loop
Tuesday, 8:12 AM CT. Dana has $140M of unallocated 2027 build capital and a board review in nine days.
Scenario timeline
Overnight autonomous work, one human conversation, immediate execution.
- 02:14Signal ScoutDetects a converging signal cluster
Overnight scan picks up 1,940 new single-family permits in Georgetown, TX, a 31% jump in net in-migration, and two AT&T Fiber conduit permits filed with Williamson County.
Signal cluster GEO-TX-114 raised to high confidence (0.91) - 02:16Signal Scout → Demand ForecasterHands the cluster to modeling
Scout passes normalized signals and geography boundaries; no human involvement required at this stage.
- 02:31Demand ForecasterSizes the opportunity
Models 14,200 new passings by 2029, 41% expected broadband penetration and 1.6 mobile lines per connected home.
$58M 5-year revenue · $11.4M mobile attach - 02:33Fiber Defense SentinelFlags a competing build
Cross-references the same permits against AT&T's regional pattern and predicts a fiber launch in Q1 2027 — roughly 11 months out — putting 6,800 existing households at risk.
$14M revenue at risk · churn exposure 4.1% - 02:40Build Economics AgentTests whether the build clears the bar
Prices the build at $780 per passing against a $950 threshold, but finds only one available construction crew in the region for Q3.
Feasible — but conflicts with the Round Rock build window - 02:41Growth OrchestratorResolves the conflict between agents
Forecaster wants the full build now; Build Economics says crews are constrained; Sentinel says the clock runs out in 11 months. The Orchestrator proposes a phased build: 6,400 passings in the competitive overlap first, remainder in Q1 2027.
Recommendation drafted with two alternatives priced - 02:44Offer DesignerStages the commercial motion
Builds an acquisition bundle for new-construction closings and a pre-emptive defensive offer — Gig + 2 mobile lines with a 24-month price lock — for the at-risk footprint.
Modeled churn impact −58% · margin dilution 3.2% - 08:12Growth Orchestrator → DanaBrings the human in, with a decision ready
A Teams card lands in Dana's morning brief: one recommendation, two priced alternatives, the assumptions behind each, and the specific approval the agent needs.
- 08:19Dana WhitfieldInterrogates the recommendation
Dana asks why Georgetown outranks Queen Creek, pushes back on the penetration assumption, and asks what happens if AT&T launches a quarter early. Each agent answers with its own evidence and re-runs its model.
- 08:26Dana WhitfieldApproves with a change
Dana accepts the phased build, moves the defensive offer 60 days earlier than the agent proposed, and holds the Round Rock decision for the regional review.
Human decision logged to Microsoft Purview with full lineage - 08:27Growth OrchestratorExecutes the approved plan
Opens the network planning work item, schedules the offer in the campaign calendar, briefs the Georgetown field team, and generates the board slide and CFO one-pager.
6 workstream items created · board deck drafted in 40 seconds
Interactive handoff transcript
Step through each handoff: what the agent said, how it changed the shared strategy, and the evidence it cited for the VP.
One living plan object every agent reads from and writes to.
- Market status
- Candidate — cluster GEO-TX-114 (0.91 confidence)
- Build plan
- Unscoped
- Revenue case
- Not modeled
- Competitive risk
- Unknown
- Commercial motion
- None staged
- Human decision
- Not requested
The 14-minute conversation
How the executive interacts with the agent — in Microsoft Teams, in plain language.
Good morning Dana. Overnight I moved Georgetown, TX from Watchlist to Priority. Recommendation: phase the build — 6,400 passings in the AT&T overlap by Q3, remainder Q1 2027 — and launch retention offers 90 days before their expected turn-up. $58M five-year opportunity, $14M of existing revenue defended. I need your approval on the phasing and the offer window.
Why Georgetown over Queen Creek? Queen Creek scored higher last month.
Queen Creek is still a stronger greenfield story, but no competitor is building there — the window stays open. Georgetown has a competitor clock: Fiber Defense Sentinel puts AT&T's launch at Q1 2027 with 0.86 confidence. Deferring Georgetown costs an estimated $9.2M in permanently lost share; deferring Queen Creek costs timing, not share.
Your 41% penetration assumption feels rich for a market where AT&T lands first.
Fair. Demand Forecaster re-ran it at 34% assuming AT&T is first to market in the overlap: the case holds at $49M with a $980 cost per passing — just above your threshold. If we hit the Q3 window and land first, it returns to 41%. That timing gap is the whole argument for phasing.
Approve the phasing. Move the defensive offer 60 days earlier and hold Round Rock for the regional review.
Done. Phased build approved, offer window moved to T-minus 150 days, Round Rock parked. I've opened the network work item, briefed the Georgetown field team, and drafted your board slide and CFO one-pager. I'll re-check AT&T permit velocity weekly and alert you if their timeline moves.
Human approval guardrails
What the agents may do alone, and where a named human must sign before anything happens.
| Gate | Trigger threshold | Approver | Rule | Mode |
|---|---|---|---|---|
| Signal & modeling | Unlimited | Autonomous | Agents may scan, score and model freely — no customer or capital impact. | Autonomous |
| Capital commitment | > $10M capex | VP, Market Strategy | Build phasing and crew allocation require a named executive approval before any work item opens. | Human approval |
| Offer & pricing | Any price-lock or >2% margin dilution | Consumer Marketing Finance | Defensive offers are staged by the agent and launched only after finance sign-off. | Human approval |
| Customer contact | Any outbound to served households | Retention Ops | No agent-generated message reaches a customer without human review of the audience and terms. | Human approval |
| Assumption override | Any model input change | Requesting executive | Humans may override any assumption; the original, the override and the rationale are all retained. | Logged override |
| Scope enforcement | Every executed action | Policy engine | Execution is diffed against the approval record — out-of-scope actions are withheld and escalated. | Policy enforced |
No capital, offer, or field commitment is executed without a named human approval.
Each number links back to its source signal, model version, and the agent that produced it.
Executives can change assumptions in plain language and the agents re-run in seconds.
Approvals, overrides and rationale are logged in Microsoft Purview with full lineage.