product / 01 · orchestration
Agent runs that survive crashes, retries, and coffee breaks.
An agent run is a long, failure-prone conversation between a model and your systems. Trellis makes each run durable. Every step is checkpointed, retries follow a policy with exponential backoff, and a run can pause for hours waiting on a human approval without holding a connection open. Then it resumes exactly where it stopped.
reading the figure
What you are looking at
-
Step 4 times out. The retry policy waits 8 seconds and tries again instead of looping.
-
A worker dies mid-run. The run moves to a healthy worker and resumes from checkpoint c4.
-
A discount over 15% needs a person. The run parks for two hours at zero cost, then finishes.
how it works
Three steps, one trace.
-
step 01
Checkpoint
Each model call and tool call is written down before it runs, so nothing happens twice.
-
step 02
Recover
If a worker or a provider fails, the run picks up from the last good checkpoint on a healthy worker.
-
step 03
Pause and resume
Approval steps park the run until a person signs off. Parked runs hold no compute and cost nothing.
capabilities
What orchestration gives you
01.1
Checkpointed steps
Every model call and tool call is saved before it executes. If a worker crashes mid-run, the run replays from the last checkpoint with no duplicate side effects and no lost context.
01.2
Retry policies, not retry loops
Declare max attempts, backoff, and which error classes are retryable per tool. Rate-limit errors back off; validation errors fail fast to the fallback branch.
01.3
Human-approval steps
Any step can require sign-off. The run parks (costing nothing), a reviewer approves in the console or by API, and the run resumes with the approval recorded in the trace.
01.4
Fan-out and joins
Spawn parallel sub-agents with a token budget each, then join on all, any, or a quorum. The runtime enforces budgets, so you are not hoping the prompt behaves.
when teams reach for it
Where it pays off first
- Runs that call slow or flaky internal APIs
- Workflows where a person must approve refunds, discounts, or writes
- Research runs that fan out into many parallel sub-agents
- Any agent you cannot afford to run twice by accident
What it costs
sample pricingThe per-run fee, nothing more
Orchestration is the one per-run meter: $0.0008 per run for the first 100k runs each month, cheaper after that. Runs parked on an approval cost nothing while they wait.
questions
Orchestration, answered plainly
What happens to a run if your service restarts?
Nothing you will notice. The run's last checkpoint is already saved, so a healthy worker picks it up and continues. The trace shows the hand-off, and no tool call runs twice.
How long can a run wait for an approval?
Up to 30 days by default. You set a timeout per approval step and choose what happens when it expires: auto-reject, escalate to someone else, or continue on a fallback branch.
Do I have to replace my queue or workflow engine?
No. Trellis runs the agent loop itself. Your existing jobs can start runs through the API and wait on the result, so the rest of your stack stays as it is.
$ trellis estimate --honest
Try orchestration on a real workflow.
Price the stack in the configurator, then bring one of your own workflows to a demo. We'll trace it live and show you this module on your data.