The Autonomous Marketing Harness

One marketer and a swarm of agents booked 33 meetings across 5,260 target accounts in a few weeks.Here is the whole engine.

Harness = ledger · coded gates · human below the bar

• Ledger: every touch written before it goes out.
• Gates as code: rules in advance, not vibes in the prompt.
• Confidence bar: above the bar it runs. Below the bar it asks.

• The brain: a frontier model, plus your playbooks, context and rules.
• Eight agents execute, one per channel, from intent to ads.
• Your team keeps the judgement calls. Sales still takes every meeting.

• 5 came through the gift and email play. The other 28 came through social outreach, pure outbound and ads.
• 8,915 people reachable out of 16,257 found, every address verified.
• It enriches, verifies, drafts and queues on its own, and hands me five decisions a day.

• Each page names the person, with their brand, audiences and ads.
• Built by metadata.io MCP tools: on brand, hosted, no agency.
• The film pipeline underneath: Kling, Fal.ai, FFmpeg, Gemini for review.

Page · MCP

Film · MCP

Creatives · Metadata

Audiences · Metadata

Their account · M1

MCP · 4 of 160 toolslanding_pagecommercialad_creativestarget_audiences

The tools

• Sense: the target list from one skill, plus G2, Bombora, web and CRM.
• Decide: a frontier model, checked by a twelve-model council.
• Act and learn: 25 skills, every touch logged, gates rewritten as code.

• Every account segmented by signal: intent, targets, closed-lost, churn.
• Every contact enriched and verified before a single touch goes out.
• Eight channels, one brain deciding who gets what, when, and how often.

• The signal picks the play. The play picks the channels and the order.
• A G2 visit earns a conversation ad, a gift and three emails inside a month.
• Five of the seven plays running today. Each dot is one touch.

• One touch a day per account: a gift, an email, a note or an ad.
• Pages and films: metadata.io MCP, on brand.
• Two outcomes are theirs: the meeting, then pipeline.

TouchWhat goes outExecuted byState

intentLights up on G2 buyer intentGBG2 · Bomboraassigned

pageTheir account, already builtMmetadata.io MCPbuilt

filmA 35-second film, on brandMmetadata.io MCPrendered

emailOne ask: see the page we builtIInstantlysent

giftA gift, then a human noteLLoop and Tiedelivered

linkedinConnection request, then a noteGGojiberryaccepted

convo adMessage ad to the buying groupMmetadata.io MCPserving

The account reaches back · two outcomes

meetingOn the calendar, traced backCCalendlybooked

pipelineTo sales, with the whole trailSSalesforceopen

• Loop and Tie gifts by seniority: $100 IC, $150 director, $250 VP.
• Seven-day expiry, ours. The default idled 213 gifts to November.
• The gift asks for nothing. The ask lives on the calendar page.

Playbook 7 of 8 · Gifting

SignalTier 1 targets and G2 intent accounts, one gift per buying group; day-0 ads through metadata.io MCP

Sequence

redeemer · 10 days

d0 emaild0 adsd1 connectd3 emaild5 messaged7 noted10 email

Stack

mmetadata.io MCPLLoop and TieCCalendlyIInstantly

APIsLoop and Tie API: POST /bulk/gifts, expires_on ours · Calendly scheduler gate · metadata.io MCP, day-0 audience

ModelsNone in the send. The council set the redeemer cadence; a survey only after a reply, never to chase.

SkillsSend a gift · Refresh gift status · Thank by gift name · Push the high-intent audience (metadata.io MCP)

CadenceLoop and Tie polled every 30 minutes as the webhook backstop · the redeemer play runs 7 touches in 10 days

Result380 sent, $46,550 · 225 unclaimed · 153 opened · 9 redeemed · 5 meetings with email · 44% of redeemers book

• metadata.io MCP: audiences, ad setup, creatives, films, attribution.
• Nine channels, six ad types, one budget: LinkedIn to ChatGPT ads.
• Above 99% it acts, below it asks: 5,418 changes in 45 days.

Playbook 8 of 8 · Ads and commercials

SignalEvery play opens with ads: always-on for targets, same-day for visitors, 30 days of high intent for redeemers

Sequence

visitor · 7 days

d0 adsd1 emaild3 connectd7 email

Stack

mmetadata.io MCPMMetadataONEffalLLinkedInGGoogle

APIsmetadata.io MCP: build_target_audiences, generate_ad_creatives, produce_commercial, campaign setup · 9 channels

ModelsAudiences, ad setup, creatives and commercials come from metadata.io MCP; a human approves anything below 99%

Skillsmetadata.io MCP: push a target audience · generate ad creatives · produce a commercial · attribute a meeting

CadenceAudiences move as plays move accounts in and out · the optimizer runs continuously, about 120 changes a day

ResultSame audience: $206 a lead on LinkedIn, $340 blended on Google · 189 G2 accounts queued for the ABM audience

• Eight playbooks. Open one and every level has its own agent.
• Ads: social, search, LLM ads and CTV, one budget, nine channels.
• Six ad types on every channel, times audiences, times offers.

• The same audience: $206 a lead on LinkedIn, $340 blended on Google.
• Confidence medium, modeled, so it asks. Above the bar it just goes.
• Three actions in one button. I read a paragraph and said yes.

5,418optimizations, one customer, 45 days. About 120 a day, no team could do that by hand.

Confidence

Above the bar
executes on its own

the bar · 99%

This one
medium, modeled. It asks.

HITLRecommendation engine

• Audit the list first: one hour lifted coverage from 9.7% to 15.8%.
• Gifting: 380 sent, 9 redeemed. The ask moved to the calendar.
• LinkedIn: 79 accepted, 0 notes sent. Running is not working.

• Targets from the metadata.io MCP audience graph, then four signals.
• No fitted model. Too few meetings to train on; the rules are code.
• One contact per company, buyer first: CMO and demand gen.

Playbook 1 of 8 · Intent and list building

Signalmetadata.io MCP audience graph for targets · G2: 6,044, 189 ours · Bombora surge 3,352 · web visitors 29 a week

Sequence

surge · 14 days

d0 adsd2 emaild4 connectd7 emaild14 email

Stack

GG2BBomborammetadata.io MCPSSalesforce

APIsmetadata.io MCP audience graph and get_deep_audience_details · G2 buyer_intent: company, state, total_activity · CSV

ModelsNone in the path. Ranking is SQL: activity, tier, then title. The council set the decay windows.

SkillsBuild the target list (metadata.io MCP audience graph) · Import a CSV · Pull G2 buyer intent · Pull Bombora surges

CadenceG2 worked the week it appears · Bombora re-dated to the audience’s own refresh · visitors within 24 hours

Result5,260 accounts · 189 on G2: 59 T1, 81 T2, 21 T3, 20 closed-lost, 2 churned · G2 play ran 5 Sep: 82 targets

• Apify finds people at $0.0015 a lead; ZeroBounce says who is real.
• Broken addresses go to SixtyFour: same person, new address.
• One hour of list audit took coverage from 9.7% to 15.8%.

Playbook 2 of 8 · Enrichment and verification

SignalEvery account from the audience graph: 25,886 leads in, ICP and domain guard, then verification before mailable

Pipeline

findICP guardverifyrepairmailable

Stack

AApifyZZeroBounceSSixtyFourCCrustdataAApollo

APIsApify: email and LinkedIn URL in one call · ZeroBounce validate · SixtyFour resolution, $0.05 a lookup

ModelsNone. Each source has a contract; answer wrongly and it is degraded, and its plays stop.

SkillsEnrich contacts · Verify every address · Re-resolve broken addresses · Audit the list

CadenceVerify before every send · preflight re-validates stale addresses · ramp only if 3-day bounce is under 2%

Result16,257 contacts · 8,915 reachable · 7,464 broken addresses, 6,457 with a LinkedIn route · spend $38.83

• metadata.io MCP builds it: their brand, audiences, ads, account.
• The film: a metadata.io MCP commercial, fal clips, cut in FFmpeg.
• A guard checks the page is live before any touch mentions it.

Playbook 3 of 8 · Landing pages

SignalEvery target account gets a page before its first email; films for the accounts worth the spend

Pipeline

briefpagefilmcheck 200email links

Stack

mmetadata.io MCPffalOOpenRouterFFFmpeg

APIsmetadata.io MCP produce_landing_page, produce_commercial, generate_ad_creatives · fal · FFmpeg · go.metadataone.com

ModelsPage copy through OpenRouter (Claude Sonnet 4.5), priced from the vendor’s usage object · Gemini reviews the film

SkillsBuild a landing page · Produce a commercial · Generate ad creatives · Check that claims match reality

CadenceBuilt ahead of day 4 of the target-account play, the touch that says “the thing we built for you”

Result493 pages return 200 · 134 with a live film.mp4 · $3 a clip, retakes capped at 2× · about $2,000 to $4,000 so far

• 17 approved templates. A missing fact refuses the send.
• Four gates in code: verified, capped, not suppressed, 48h cooldown.
• Ledger first, then Instantly: a crash cannot double-send.

Playbook 4 of 8 · Outbound email

SignalA play owes a touch today: G2 intent, surge, target account, closed-lost, churned or visitor

Sequence

target account · 30 days

d0 adsd0 connectd4 emaild11 messaged18 emaild30 email

Stack

mmetadata.io MCPIInstantlyZZeroBounceGGmail

APIsInstantly API v2: leads, campaigns, accounts · ZeroBounce · metadata.io MCP for the day-0 audience · Postgres ledger

ModelsNone in the send path. The council set the cadence: one touch per channel per 48 hours after the first 72.

SkillsPlan the day’s sends · Send the queue · Preflight, 14 checks, arm switch · Bounces · Canary · Unstick

Cadence50 a day now · 100 at day 5 · 150 at day 12 · 250 at day 21, each step only if 3-day bounce stays under 2%

ResultThree inboxes, 250 a day usable · G2 play: 82 targets, 29 held to Tuesday · cold closed-lost email: 0 of 173 replied

• Connect first, no pitch. The message goes out on accept only.
• 6,457 people with a dead email still have a LinkedIn route.
• Running is not working: no message step, and nothing said so.

Playbook 5 of 8 · LinkedIn outreach

SignalContacts with a LinkedIn URL inside an active play; closed-lost and target accounts connect on day 0

Sequence

closed-lost · 21 days

d0 connectd2 emaild5 messaged9 emaild14 giftd21 email

Stack

GGojiberryAApifyCCalendly

APIsGojiberry v1: campaigns, unibox send-message · Apify for the LinkedIn URL · Calendly as the meeting sensor

ModelsNone in the send. Copy from approved templates; the council put the cooldown after a personal email at 7 days.

SkillsLinkedIn connect and note · Message on accept · Sync accepted connections to the ledger

CadenceConnect day 0 · message on accept, day 5 or 11 by play · Gojiberry state polled every 60 minutes, stale at 120

Result79 accepted · 0 messaged before the fix · the engine now reads accepted and messaged as two different states

• One chapter per part of the engine; every number from the ledger.
• Chapter 0 goes up this week. Chapter 1 is the gifting play.
• Engagers become a list source for the competitive takeout play.

Playbook 6 of 8 · LinkedIn posts

SignalWho engages with each chapter, matched against the target list and the plays already running

Pipeline

writepublishengagersmatchenter a play

Stack

LLinkedInAApifyPPostgreSQL

APIsLinkedIn · Apify, LinkedIn social signals only · the engine’s own tables for every number quoted

ModelsDrafting is assisted; every figure in a chapter is a query against the engine database, not a recollection.

SkillsAdd a chapter · Pull engagers · Match engagers to accounts

Cadence15 chapters, one at a time, starting the week after the keynote

Result15 written · 0 published · build_in_public and competitive_takeout plays defined, waiting on the engager pull

• A page, a film, an email, a gift, a conversation ad, a LinkedIn note.
• Drafted from approved copy, checked against the person, logged first.
• I approve. I do not write.

The page · live

The film · 44 s

Email · real send

Gift · Loop and Tie

GGil Allouche inSponsored

Dana, we built [Account] a working account. Yours to look at.

See the page

Convo ad · mock

GGil Allouche inFounder and CEO, Metadata

Dana, we built [Account] the ads we would run. Worth a look?

IgnoreAccept

LinkedIn · mock

• Metadata: creatives, films, audiences, attribution.
• Twelve vendors execute: intent, data, email, gifts.
• One model decides. Twelve check it. The gates hold.

VendorJobWhat it does for us

MMetadataads, pages, filmsCreatives, audiences, attribution, campaign execution, optimization

LLoop and Tiegifts $100 to $250380 sent · 153 opened · 9 redeemed · 5 meetings with the email play

IInstantlyoutbound emailSends the sequences from warmed inboxes, 50 a day

BBomboraintentSurging accounts on our own list, 3,352 today

GG2intentWho is reading our G2 profile: 189 of 6,044 are ours

AApolloenrichmentPeople and emails, one credit per match

SSixtyFourenrichmentRe-resolves broken addresses, $0.05 a lookup

CCrustdataenrichmentFinds the people at an account

AApifyenrichmentLeads with the email attached, $0.0015 a lead

ZZeroBounceverificationVerifies every address before a send

GGojiberryLinkedIn outreachConnection requests and notes

CCalendlymeetingsWhich meetings, with whom, from where

EExaresearchWeb research on people and companies

• One specialist per channel, each with their own tool and their own numbers.
• Four owners touch the same account and never compare notes.
• Sales executes the last mile, one more handoff on every play.

• LinkedIn, Meta, Instagram, Reddit, X, Google, Microsoft, ChatGPT ads.
• Only target accounts, and the personas inside them, ever see an ad.
• Spend to closed-won in one journey, natively or through the MCP server.

• 15 chapters written. Chapter 0 goes up this week.
• Every number in every chapter is measured in the engine database.
• The gifting, the outbound, the intent, the pages, the ads, the numbers.