Slack Engineering

Slack's four-phase path from managed SageMaker inference to multi-vendor LLM orchestration, including the escrow-VPC arrangement that kept customer data private and model weights opaque, and the scaling-latency and capacity-planning taxes that pushed them off a single provider.

Slack put Playwright MCP, the Playwright CLI and agent-authored tests through 200+ runs to see where agentic E2E actually earns its place. The finding: agents reach the same outcome by different routes, which is coverage and flakiness at once. A framing for layering both.

Slack's security team on keeping a Director/Expert/Critic agent team coherent across investigations with no fixed round limit — the techniques they use to balance continuity against creativity as accumulated message history pushes against the context window.

How Slack's security team turned a 300-word prompt into an alert-investigation agent: MCP as the controlled data-access boundary, a coding-agent CLI as the execution environment, and an honest account of how unreliable the first version was. First in a series.