The fleet

20 specialists. No agent does two jobs.

A0 Conductor — a single orchestrator agent — was scoped and deliberately not built: a fixed pipeline plus a reactive cost cutter does that job without adding a layer that can itself hallucinate a decision. What's below is what actually runs.

Layer 00

CONTROL

Independent verification and spend control — the checkpoint layer that watches the rest of the line.

Auditor

Independent verifier. Produces nothing — approves or rejects with a reason and a confidence score.

Cost Guardian

Pure code, no LLM. Routes to the right model and cuts off a job before it blows the budget.

Layer 01

ACQUISITION

How Agenclaw finds and grows its own customer base — the same engine it sells, running on itself.

Prospector

Surfaces evidence-backed prospects from hiring signals, funding, weak SEO and competitor churn.

Outreach

Writes signal-specific personalized cold email and classifies replies.

Growth-PSEO

Builds agenclaw's own programmatic pages — this site runs on the same engine it sells.

A20

Blog Author

Writes Agenclaw's own editorial blog — a separate role from A5's templated own-site pages.

Layer 02

LEAD GENERATION

Sources, scores, and verifies your contacts — three separate checkpoints, three separate agents.

ICP Interpreter

Turns your two-minute form into a machine-readable ICP spec.

Sourcer

Finds matching companies and people — every row backed by evidence.

Enricher

Finds and SMTP-verifies work emails. No verification, no delivery.

Qualifier

Scores fit 0–100 with a reason — sub-criteria always sum to the score.

A10

Compliance Sentinel

Pure code, no LLM. Blocks role accounts, suppressed contacts, GDPR/CAN-SPAM/KVKK violations.

Layer 03

PROGRAMMATIC SEO

Maps intent, gathers evidence, writes your pages — in that order, never reversed.

A11

Keyword Miner

Mines long-tail targets from seed keywords — volume and difficulty are never invented.

A12

Cluster & Cannibalization

Clusters keywords and flags cannibalization via SERP overlap.

A13

Content Architect

Builds the brief and internal link map — no orphan pages.

A14

Writer

Writes the page. Every number is tied back to a source.

A15

Editor & Fact-Check

Fact-checks, screens for unsourced numbers, watches for content decay.

Layer 04

DELIVERY & OVERSIGHT

Gets finished work to you, and keeps a record of what happened.

A16

SEO Technician

Mostly code. JSON-LD, meta, canonical, sitemap, broken-link checks.

A17

Publisher

Publishes straight into WordPress, Webflow, Ghost, Shopify, Sanity, Contentful, or a webhook.

A18

Reporter

Builds the weekly customer performance report.

A19

Dead-Letter Triage

Classifies failures as transient, config, or quality — resolves most without a human.

The real question

How do you stop twenty agents from hallucinating?

You don't trust any single agent's output. You make every claim traceable, and you put a separate agent in charge of checking, not producing.

Separation of production and judgementA9 Qualifier scores a lead. A1 Auditor checks the score and the rationale independently. A14 Writer drafts a page; A15 Editor & Fact-Check verifies its claims. The agent that made something never approves its own work.

Evidence is mandatory, not advisoryA claim without a traceable source doesn't get "probably fine" treatment. It's held for rewrite or spiked at the gate — the same rubric described on the How it works page.

Confidence thresholds, not blind outputAgent configurations carry a minimum confidence requirement. Output below that bar is routed for review rather than shipped on the assumption it's probably right.

A human decision point, alwaysNothing publishes or sends without an approval step — either your standing settings or an explicit action in your dashboard. The machine executes; it doesn't have the last word.

Cost ceilings enforced before spend, not afterA2 Cost Guardian checks a run against your cap before it starts and mid-run. A run that would breach it is refused up front, not explained in an invoice afterward.