12 个仓库给 GPT-6 智能体做崩溃恢复
用 Restate、DBOS、Windmill、NATS、PostgreSQL 等 12 个仓库，给 GPT-6 智能体搭一套跨崩溃、重启、断连的持久化执行架构。
AI agents are getting f…cking impossible to stop 12 GitHub repos for building a GPT 6 system that keeps working across crashes, restarts, failed tools, and disconnected sessions 01 Restate ▸ https://github.com/restatedev/restate → durable execution for agents and distributed workflows MAKE EXECUTION SURVIVE 02 DBOS ▸ https://github.com/dbos-inc/dbos-transact-py → recover Python workflows after interruptions 03 Windmill ▸ https://github.com/windmill-labs/windmill → run scripts, workflows, and background jobs 04 http://Trigger.dev ▸ https://github.com/triggerdotdev/trigger.dev → durable background tasks with retries KEEP THE WORKERS MOVING 05 NATS ▸ https://github.com/nats-io/nats-server → messaging between distributed workers 06 Celery ▸ https://github.com/celery/celery → distribute tasks across worker processes 07 BullMQ ▸ https://github.com/taskforcesh/bullmq → Redis backed job queues with retries SAVE THE STATE 08 Valkey ▸ https://github.com/valkey-io/valkey → fast state storage and coordination 09 PostgreSQL ▸ https://github.com/postgres/postgres → persist jobs, checkpoints, and execution history 10 MinIO ▸ https://github.com/minio/minio → store artifacts and intermediate outputs RECOVER AND VERIFY 11 OpenTelemetry ▸ https://github.com/open-telemetry/opentelemetry-collector → collect traces and telemetry across services 12 Argo Workflows ▸ https://github.com/argoproj/argo-workflows → coordinate container based workflows on Kubernetes the architecture: task → queue → execute → checkpoint → failure → restore → retry → verify → continue 3 builds worth bookmarking: Python agents: DBOS → Celery → PostgreSQL distributed workers: Restate → NATS → Valkey production workflows: Windmill → Argo Workflows → OpenTelemetry the biggest change isn’t making GPT 6 Astra run more tasks it’s making the surrounding infrastructure survive when those tasks go wrong a crashed process shouldn’t mean a lost mission the next generation of agents won’t need perfect runs they’ll need systems that know how to continue ⭣
这条是引用着另一条写的：x.com/lummox_eth/status/2100288151897981255