ENGINEERING
How we cut agent inference latency to 82ms
A look inside the scheduling, caching, and routing work that took our median agent response time below the 100ms...
PRODUCT
Observability that speaks "agent"
Traces, token budgets, and replay — built for non-deterministic runs.
SECURITY
Scoped credentials for autonomous tools
Per-tool, per-run secrets that expire the moment a task completes.
ENGINEERING
Shipping reliable tool-calls at scale
How we cut tool-call failures by 9× with typed schemas and retries.
ENGINEERING
Streaming responses without the jank
Backpressure, flush cadence, and the 41ms p95 we hold in production.
ENGINEERING
Tracing across multi-agent systems
Following a single request as it fans out across a swarm of cooperating agents.
PRODUCT
Vector memory, explained for builders
A practical mental model for giving agents durable, queryable memory.
TUTORIALS
Zero-downtime canary deploys for agents
Roll new agent versions to 5% of traffic, watch p95, then promote.