Infrastructure & compute

Status: complete (published 2026-08-19). Scope finalized at Batch A kickoff with maintainer POV; see brief.md, findings.md, vendors.md, sources.md.

Scope inventory (v0)

Current estate: cloud landing zones, multi-cloud and region strategy, Kubernetes, serverless, IaC and GitOps provisioning automation, environment strategy (dev/test/prod parity for agents)
Agent execution: sandboxes and isolation (microVMs, containers), browser and computer-use infrastructure, VDI where agents act on desktops, ephemeral cattle environments with lazy provisioning, cold-start and time-to-first-token economics
Model serving: GPU and accelerator capacity planning, inference serving for self-hosted and open-weight models, edge inference for OT and remote sites
Networking for agents: egress control and allowlists, forward proxies, private connectivity to SaaS and model providers, service mesh
Identity and resilience: workload identity for agent compute, infra-level secrets and short-lived credentials, checkpoint and session persistence, crash recovery, DR, quotas and noisy-neighbor controls
Cost: resource tagging and showback, autoscaling policy, spot and reserved capacity strategy

Self-hosted GPU estates vs metered APIs; dedicated agent-sandbox products vs hardened containers. Each track proposes its final list at kickoff.

brief.md, findings.md, vendors.md, sources.md are created from ../_TEMPLATE/ at kickoff.

Source: research/R01-infrastructure/README.md in the evidence repository behind this site.