Infrastructure as Context

niels-bantilan's design principle: agents need visibility into their own runtime to self-heal.

"A key thing that we found that's quite important when you're building sort of variable-compute agents that require variable compute in the inner loop. Providing agent self-healing utilities… Self-healing is a little buzzwordy, but basically: how do you recover from failure in the inner loop?"

Concretely: when an agent's inner loop hits a timeout, an OOM, or a missing dependency, it should be able to reason about the infra (not just the semantic error) — query its own compute budget, spin up a sandbox, retry with more RAM.

Why this matters

Most agent harnesses treat infra as invisible plumbing. Bantilan's argument is that self-healing is impossible without exposing that layer to the agent. This is a non-obvious corollary of durable-observable-debuggable-agents — durability is table stakes; self-healing requires durability plus infra-awareness.