Layer

The problem

A salesperson’s day is spread across a CRM, sales engagement and call tools, Slack, email, a calendar and a string of calls. Layer puts an agent across all of it. The agent organises the whole working day: it suggests which accounts to move and what to do next, coaches the rep over time, prepares them before a meeting, listens during the call and offers live suggestions, and handles the follow-up afterwards. The rep sells; the agent carries the context.

What I worked on

• Core agent loop. The run loop, tool execution and how a run recovers from failures mid-task.
• Context. What the agent sees about an account, a meeting or a thread, and how that stays small enough to be useful.
• Integrations. Salesforce, HubSpot, Salesloft, Gong, Slack, Gmail and calendar connectors, including auth, sync and the write paths the agent was trusted with.
• Billing. Integrated billing and charging for units of work across the agent, so usage turns into something a finance team can invoice.
• Briefs and post-call automation. Pre-meeting briefs, and the CRM updates, notes and follow-ups generated after a call.

What it taught me about agent products

An agent that writes to a CRM is trusted with a customer’s source of truth, and an agent that speaks during a live call is trusted with the customer relationship itself. Most of the engineering effort went into making that trust reasonable: bounded tools, reviewable actions, latency budgets that hold up mid-conversation, and context that is deliberately curated rather than everything the integration can fetch.

It is also where I learned that a production agent loop is mostly plumbing done carefully. The model is the smallest part of the system. Recovery, idempotent writes, metering and the shape of the context are what decide whether the product survives contact with a real sales team.