PLATFORM · AGENTS

AI that does one job extremely well.

Purpose-built agents for narrow operational roles. Not a chatbot with your logo on it. Something that decides and acts inside your workflow.

Start a project → Browse the catalog
01 · AGENT CATALOG
RESEARCH
runs/day · ~140
Scout
Reads everything, surfaces what matters.
Inputs
  • + Source feeds
  • + Competitor sites
  • + Regulatory filings
  • + Your knowledge base
Decides
  • + Is this new?
  • + Is it relevant?
  • + Who owns it on our side?
Escalates on New regulatory risk · competitor pricing moves · unverified claim.
SALES INTEL
runs/day · ~80
Qualifier
Inbound leads enriched and scored before anyone picks up.
Inputs
  • + Form submissions
  • + Firmographics
  • + Intent signals
  • + CRM history
Decides
  • + Fit score
  • + Routing
  • + First-touch draft
Escalates on Enterprise fit · existing account · hot signal but missing contact.
MONITORING
runs/min · ~1
Sentinel
Watches the stack and speaks up only when it matters.
Inputs
  • + Logs
  • + Uptime checks
  • + Customer sentiment streams
  • + Billing anomalies
Decides
  • + Page on-call or not
  • + Which team
  • + Severity
Escalates on Novel incident shape · cross-service correlation · customer-visible.
DOC OPS
runs/hour · ~40
Clerk
Intake, extract, classify, file, cross-reference.
Inputs
  • + Incoming PDFs / emails
  • + Document taxonomy
  • + Record templates
Decides
  • + Extracted values
  • + Record type
  • + Where it filed
Escalates on Low-confidence extraction · missing fields · conflicting duplicate.
CUSTOMER OPS
runs/day · ~400
Concierge
First-touch triage that drafts the response a human will send.
Inputs
  • + Ticket text
  • + Account context
  • + Past tickets
  • + Product docs
Decides
  • + Intent class
  • + Queue
  • + Draft reply
Escalates on Churn-risk language · legal/contract content · safety concern.
02 · ANATOMY OF AN AGENT

Five layers, in plain terms.

Every agent we ship has the same five-layer build. Goal, context, tools, judgement, governance. No layer is optional.

01
GOAL
A narrow, measurable job

Not "be helpful." Something like: "qualify inbound leads with ≥0.85 precision on fit."

Written spec · monthly SLO
02
CONTEXT
The world the agent sees

Your data, your docs, and current state. Scoped, not the open internet.

pgvector · Postgres · S3
03
TOOLS
What it can do

A bounded set of typed actions. Fetch a record, draft an email, tag a ticket, escalate. Five to ten per agent, no shell.

Anthropic tool use · JSON schema
04
JUDGEMENT
How it decides

Haiku for extraction. Sonnet for routing. Local models for high-frequency tags. Right model, right task.

Claude · GPT · local
05
GOVERNANCE
Who checks the work

Every action logged as JSONL: agent, run_id, inputs, decision, confidence. Reviewable, replayable, tunable.

JSONL · 90-day · replayable
03 · GUARDRAILS
BR · 01
Bounded tools

Agents only call tools you approved. Concierge ships with fetch_account, draft_reply, escalate. No shell, no arbitrary HTTP.

BR · 02
Confidence thresholds

Below threshold, it escalates. Clerk extracts at ≥0.92, Sentinel pages on-call at ≥0.85. You set the bar.

BR · 03
Human-in-loop

High-impact actions queue for one-click approval. Outbound mail, payment, write-back to CRM. Configurable per tool.

BR · 04
Full audit trail

Every decision logged as JSONL: agent, run_id, tool, inputs, output, confidence. Replayable. 90-day retention by default.

GET IN TOUCH

Describe the role. We'll scope the agent.
Start with one job.

Most deployments start with a single narrow agent and expand once the ROI is obvious.

Start a project → Browse the catalog