Platform
How Layap runs
An AI agent orchestration platform built to run our own products in the market. Agents propose. You approve. The ledger records every action.
Architecture
Layap runs as a control plane plus a fleet of cron-triggered and event-triggered agents. Every agent bootstraps from a manifest: job types, required tools, risk tier, and pipeline steps. Outcome events flow to MongoDB. Approvals gate every external action.
Approval system
The Approval Inbox is the operator seat. Agents file tickets for outreach sends, account creates, retention emails, social posts, and reply drafts. You edit, approve, or reject in one click. Nothing sends unverified. Nothing publishes without review.
Playbooks
Each agent resolves an active playbook at boot — ICP scoring weights, email voice, objection handling, calendar links. Playbooks version over time as the Human Edit Miner surfaces operator edits as insight tickets.
Connectors
L1 read-only metrics. L2 account read/create. L3 workflow triggers. Connectors scope tenant isolation — agents never cross network boundaries.
Cost metering
Every LLM call is metered to the cent per tenant, per agent, per job type. The Weekly Reporter includes platform LLM spend alongside funnel metrics.
Learning loop
Human Edit Miner clusters approval diffs. Scoring Calibrator recomputes ICP signal lift monthly. Eval Runner runs golden-set fixtures weekly. Analyst files 3–7 insight tickets every Monday. The system compounds operator instincts.