Agent Infrastructure: Memory, Sessions, Orchestration, Cost, Tests

MinIO pitches persistent memory for agents with work to finish. MinIO launches AIStor Memory to give agents persistent, enterprise-controlled memory that unifies storage, workspaces, and secrets under customer governance. Outcome engineers should treat durable, governed context stores as first-class infrastructure—design your context layers to support provenance, access controls, and long-lived state (Principle 06).

Mitchellh starts a new company: Superlogical. Superlogical builds a durable multiplexer session that unifies local, remote, CI, agents, and production to preserve context, actions, and history across work. This reduces handoff friction and makes agent-driven workflows reproducible and debuggable, so prioritize session durability and replayability in your tooling (Principles 06 & 11).

Enterprise AI agents can’t talk to each other, can’t be trusted with permissions, and can’t be audited — 5 startups are already fixing that. Startups are shipping orchestration, governance, and audit layers so enterprise agents can communicate, get trusted permissions, and be audited. If you build agent stacks, adopt composable orchestration and observability primitives now—inter-agent contracts, permissioned capabilities, and auditable traces are becoming the baseline (Principle 09).

Launch HN: Tokenless (YC S26) — Automatic model switching to save money. Tokenless halves inference costs by fan-out routing requests to multiple models, selecting the optimal one and cancelling redundant runs. Cost-aware model routing changes architecting trade-offs for latency, accuracy, and budget—bake model selection and cancellation semantics into your orchestration fabric (Principles 09 & 12).

Shipping an MCP Test Agent — The Boring Parts Nobody Demos. The piece spells out runbooks, strict agent-to-agent contracts, and provenance tracing as essentials to make unattended agentic test pipelines safe, debuggable, and production-ready. Treat agent testing as infrastructure: codify runbooks, create deterministic contracts, and convert incidents into reproducible tests so evaluations gate deployments (Principles 06, 13 & 14).