Work starts itself.
Wake an agent when CI fails, an issue opens, or a webhook arrives. Replace repeated checking with work that starts at the right moment.
sys_subscribe Open source · Apache 2.0 · Alpha
Your agents.
Already on it.
Even when you’re off.
WakeVane is an open-source harness for long-running AI agents. Turn events into action, keep work moving in the cloud, and see what every session costs.
Built on Databricks’ Omnigent. Designed to earn its keep.
THE EVENT-DRIVEN WORKDAY● ON CALL
↳ Issue opened
× CI failed
↑ Deployed
WakeVaneWAKE → ROUTE → GOVERN
TriageUnderstand
FixProve the change
WatchVerify the deploy
☁ CLOUD SANDBOXESPOLICIES + SPEND CAPS
→ event received · check_run.failed
→ foreman awake · fixer dispatched
✓ your next step · review the evidence
ILLUSTRATIVE WORKFLOW / NOT LIVE ACTIVITY
One server.
Your choice of agent.
Your choice of agent.
Claude CodeCodexCursorOpenCodePi+ your own
01 / LESS SUPERVISION. MORE MOMENTUM.
Let models do the thinking.
Let the harness do the rest.
Your time belongs in decisions. WakeVane handles the triggers, runtime, and boundaries that keep agent work moving.
Close the laptop.
Run long tasks in cloud sandboxes. Pick up the same session from your terminal, browser, or phone when you’re ready.
Persistent cloud runtimeKeep the controls.
Set policies to allow, deny, or ask before acting. Use approvals and spend caps to define what agents can do on their own.
Your rules. Your budget.Know what it costs.
Inspect cache hits, cost per session, and model mix. Start with the most expensive sessions and find where to improve.
Measured, not assumed.02 / THE SOFTWARE FACTORY
From a red build
to a reviewed fix.
A foreman coordinates the work. A fixer proves the change. A watcher follows the deploy. You stay in charge of the merge.
FACTORY / WALKTHROUGH01 OF 04
EVENT → TRIAGE
The right work, at the right time.
→ Received: check_run.failed → Foreman: read logs and classify the failure → Route: sandboxed fixer No one had to open a terminal.
● Example flow · human approval before merge
Illustrative sequence. Results depend on the task and configuration.
03 / PUT IT TO WORK
Less checking in.
More moving forward.
For builders creating autonomous colleagues, platform teams managing spend, and engineers evolving their harness.
CI and bug triage
Turn failed builds and incoming issues into diagnoses or proposed fixes. Review the result instead of starting every investigation yourself.
After-deploy monitoring
Keep a watcher on the rollout. Check health after a deploy and bring regressions back to the team with context.
Multiple agents, one goal
Use the Polly example to split refactors or migrations across agents and have one review another’s work.
Security and docs on repeat
Connect review and documentation agents to repository events. Make security checks and release notes part of the workflow.
Long tasks, fewer interruptions
Start a multi-hour task in the cloud, step away, and return to the session from another device.
A budget you can inspect
Compare harnesses and models, investigate costly sessions, and cap spend before unattended work runs away.
THE HARNESS THAT EARNS ITS KEEP
Measures itself.
Deletes its own scaffolding.
Every registered patch over a model weakness—a “shim”—has a reason, a review-by release, and a log of when it fires. A janitor agent opens PRs to remove shims that stop being useful.
As models improve, the harness can get smaller. Use the efficiency scorecard to see where the cost goes.
Read about the shim registry ↗$ omnigent usage
--efficiency
--efficiency
- Cache hit rate
- Inspect reuse
- Cost per session
- Find expensive work
- Model mix
- Match model to task
- Worst sessions first
- Prioritize improvements
- By harness + model
- Compare the breakdown
Scorecard capabilities shown here.
Run it on your usage for real measurements.
04 / A FEW GOOD QUESTIONS
A harness with
a clear purpose.
What exactly is WakeVane?
WakeVane is an open-source distribution of Databricks’ Apache-2.0 licensed Omnigent. It adds event subscriptions, a shim registry, and an efficiency scorecard to a shared server for coding agents. It is currently alpha.
What comes from Omnigent?
The multi-agent runtime, cloud sandbox support across 12 providers, device-spanning sessions, policies, approvals, and spend controls. WakeVane keeps the upstream omnigent CLI name to make compatibility and ongoing merges easier.
When is it worth using?
When tasks should start on an event, continue without your laptop, or involve several agents. For a quick interactive coding task, a standalone agent is often simpler.
Will it make my agents cheaper?
It gives you tools to measure and manage cost, not a guaranteed saving. Run a focused pilot, inspect actual sessions, and compare cost alongside the quality of the completed work.
Can I keep humans in the loop?
Yes. Configure policies to allow, deny, or ask before actions, with approvals and spend caps. The software factory example asks a human before merging a proposed fix.
START SMALL. MEASURE THE DIFFERENCE.
Give one repo an agent on call.
Try the factory on CI-failure events for a week. Review the fixes, inspect the efficiency scorecard, and decide whether it earns its place.