Agents that run on Orka, and the tests that gate changes to them.
| Agent | What it does | Runtime | Status |
|---|---|---|---|
| a2a-assistant | A no-tools assistant exercised over A2A | native | tested |
| coordinator | Runs the parent on Azure and delegates one level deep to the catalogue/promotion-pinned local hello child |
native | tested, trial/production verify green |
| dependabot-repair | Fixes failing CI on dependency update PRs | claude | tested, ran on real PRs |
| hello | The smallest possible agent. Start here. | native | tested |
| release-notes | Drafts release notes from a list of merged changes | claude | tested in simulation |
Render the bundle, install it, and submit an explicit Task for hello, the
catalogue's front door. You need a Ready native Provider, envsubst, and
CONTEXT/KUBECONFIG set.
tools/render agents/hello trial --output /tmp/hello.json
kubectl --context "$CONTEXT" --kubeconfig "$KUBECONFIG" apply -f /tmp/hello.json
NAMESPACE=orka-system PROVIDER=hello envsubst '${NAMESPACE} ${PROVIDER}' < agents/hello/examples/task.yaml | kubectl --context "$CONTEXT" --kubeconfig "$KUBECONFIG" create -f -See the hello README for its Provider requirements and how to read the Task's result.
Set AGENT to the directory name, then render and check the change.
tools/render "agents/$AGENT" trial --output "/tmp/$AGENT.json"
tools/verify "agents/$AGENT" trialIf behavior changed, run the tests listed in that agent's README. They write a
small JSON receipt under eval/receipts/<version-digest>/. Commit that receipt
with the edit. The gate turns red when the receipt is missing, belongs to a
different version, or records a failed test. The full guide explains the
loop.
A version is a digest of the prompt, Orka resources, dependency lock, acceptance rules, memory baseline, and environment overlay. Test receipts are tied to that digest. CI checks them offline and never talks to a cluster.
Deploy applies the rendered bundle and records what the cluster reads back.
Rollback uses git revert, so the earlier digest and its receipts return. Then
we deploy that bundle and run tools/rollback-verify. Read the full
mechanism.
agents/
a2a-assistant/
coordinator/
dependabot-repair/
hello/
release-notes/
docs/
tools/
tests/
These agents run on our own repositories so we can try running agents on AKS; they are not the team's operating process. Each agent runs only when asked, holds no credentials in its runtime, and publishes through a separate trusted component, so the prompts are public.