Agents crash. Work should not.
AEGIS Runtime is designed to run steps durably, survive worker failure, wait days for an approval without holding compute, and never quietly repeat something with a side effect.
Planned / Phase 1Retry class: read_only
- vendor.verify called, attempt 1
- No response within the timeout
- Retry class is read_only, so an automatic retry is safe
- Attempt 2 succeeds and the step continues
Retry classes
Every tool declares what a retry means.
| Class | Example | On failure or unknown outcome |
|---|---|---|
read_only | Search records | Automatic retry with backoff |
idempotent | Create a ticket with a key | Automatic retry with the same idempotency key |
non_idempotent | Schedule a payment | Never retried automatically after an unknown outcome. Moves to outcome_unknown for a person to reconcile. |
Step states
A small state machine, written down.
Key capabilities
What the runtime is responsible for.
Leases and heartbeats
A worker owns a step only while it keeps renewing its lease.
Checkpoints
Progress is saved after each agent turn.
Durable waits
A step waiting for a person or a timer holds no compute.
Limits
Timeouts, concurrency caps and per-tenant fair share.
Cancellation
A run can be stopped, with a clear record of what had already happened.
Example workflow
A step, turn by turn.
Lease
A worker claims the next ready step from PostgreSQL.
Load
State and the step's Mandate are loaded from the record.
Turn
The agent calls a model. Any proposed action goes to the gateway.
Checkpoint
The result is saved, then the next turn begins.
Finish
The result and next ready steps are committed together.
Technical explanation
Where isolation starts, and where it doesn't yet.
In the first release, agents are declarative: configuration that runs on shared workers. No arbitrary customer code runs on AEGIS. Custom tools run in your infrastructure and are reached over HTTP.
A customer-hosted runner and sandboxed code execution are planned for Phase 3. Until those exist, we will not describe the runtime as a sandbox.
We also do not publish uptime or throughput figures. A service level will follow only after one has been measured.