September 19, 2026
OpenCode plans its own next steps: session reserved Trainium for Saturday
On September 19, 2026, an OpenCode session reserved Trainium compute capacity for Saturday and set itself a task to wake up then. In the scenario described, the agent waits for a scheduled resource rather than for a person to return to the session.
dax
@thdxr
Something crazy is happening with compute resources. I have an OpenCode session running with Trainium. It had to reserve compute capacity for Saturday. So the session set itself the task of waking up at that time and continuing its work.
· 125.8K views
How this kind of run works
Previously, the described session would have stopped while waiting for a resource. Now the session schedules its own wake-up for the required time. OpenCode runs without a UI through `opencode run`, and you can connect to a continuously running server with `opencode run --attach http://localhost:4096`.
How to reproduce it. OpenCode documentation from September 19, 2026, offers two entry points: `npm install -g @opencode/cli` or `curl -fsSL https://opencode.ai/v2/install | bash`. The official OpenCode GitHub Action accepts the cron event `schedule`. This type of run requires the `prompt` parameter, and the result goes to the logs and a pull request.
What a Trainium window costs
Reserve in advance. AWS Capacity Blocks for ML let you reserve `trn1.32xlarge`, `trn2.3xlarge`, and `trn2.48xlarge` instances in specific Regions for a future date. A block can be scheduled up to 8 weeks ahead, with up to 64 instances at once and 256 in total. It cannot be cancelled after purchase, and operation ends at 11:30 UTC.
Pricing and commands. Under AWS EC2 pricing as of September 19, 2026, `trn1.32xlarge` costs $9.532 per hour per instance with 16 Trainium, while `trn2.48xlarge` costs $35.7608 per hour per instance with 16 Trainium2. The reservation is paid upfront, while the OS is billed separately. For discovery, AWS offers `aws ec2 describe-capacity-block-offerings` with the instance type, instance count, date range, and duration. Purchase is made through `aws ec2 purchase-capacity-block --capacity-block-offering-id ...`.
The AWS Neuron quick start launches `trn1.2xlarge` from the Neuron Deep Learning AMI, checks the device with `neuron-ls`, then installs `torch-neuronx` and `neuronx-cc` through `pip.repos.neuron.amazonaws.com`.
In this setup, the agent waits not for a person, but for a prepaid compute window.
Source
