The Next Scaling Problem

Summary

Long-form engineering essay on scaling cloud agents by pulling the runtime out of the sandbox and rebuilding the system around durable agent work. Core principles: (1) A computer should be something the agent calls, not somewhere the agent lives -- the agent is a stable shared service while sandboxes are disposable execution resources with their own lifecycle (models, memory, files, credentials, tools each get one). (2) Write ahead of execution -- a WAL-style protocol: record a transition before performing the operation it authorizes, with stable identities as idempotency keys, so a replacement runtime can reconstruct a thread from committed facts after a failure. (3) Durable delivery via a queue that outlives producers. Describes the Tetral architecture: Gateway (provider integration, stateless), Bridge (PostgreSQL commits with ownership/order checks), Queue (ordering, leases, retries, dead-lettering), Sandbox Service (computer lifecycle), Public API + Event Stream; the runtime itself is replaceable compute running a pure reducer with no DB/network I/O. Storage model: session_events (ordered event log), session_messages (model-context projection), session_bridge_operations (idempotency receipts); compaction adapted from OpenCode. Contrasts with Cursor's Temporal-based durable-execution approach.

Keywords

agents · cloud infrastructure · sandbox · durable execution · Tetral

Topics

agentic systems, cloud infrastructure, durable execution

Research notes

• Discovery: recommended by Yifan Xu (@yifanxu_ephai, CS @ UofT; author of LayerFS and Ephemera) on 2026-09-29: https://x.com/yifanxu_ephai/status/2104867172527087884
• Author: Yang Li (@YangLi_leo), founder of Tetral (previously built Anoma, a cloud agent product on E2B).
• The tweet highlights four chapters: "A computer should be something the agent calls, not somewhere the agent lives", "Write ahead of execution", "A durable participant", "Scaling the system around intelligence"; and explains why agents should live outside the sandbox with compute/storage separated, plus how that architecture enables multi-agent workspaces.