9. What's Next
The pattern is complete and runnable today. These are the directions that make it stronger, roughly in priority order. Several are good first contributions, see Contributing.
flowchart TD
now[shipped today<br/>complete pattern, laptop to prod] --> near[Near term]
now --> mid[Mid term]
now --> long[Longer term]
near --> n1[upstream proposals filed]
near --> n2[maintainer diversity<br/>and public adopters]
mid --> m1[OCM cluster-proxy transport]
mid --> m2[authenticated HTTP transport]
mid --> m3[off-box approval signer]
long --> l1[more chaos classes:<br/>node pressure, partitions, noisy neighbors]
long --> l2[Placement-aware fleet actions]
long --> l3[policy-pack presets per industry]
Near term #
- The remaining upstream proposals. The Kyverno one is filed:
kyverno/policies#1534
proposes the ManifestWork-envelope pattern to the policy catalog, since nothing
there demonstrates validating workloads embedded inside another CR. Two drafts
remain in
docs/upstream-notes.md: long-running operations in MCP, and richer ManifestWork failure feedback in OCM. - Maintainer diversity and public adopters. These gate CNCF Sandbox and are the only steps that cannot be engineered - see CNCF readiness.
Shipped: a recorded end-to-end demo in the README, and published evaluation results across agents from different vendors with the failures called out - see Evaluation.
Mid term #
- OCM cluster-proxy transport so the server host never holds spoke credentials directly.
- An authenticated HTTP transport (SSO/OIDC) so the in-cluster Deployment can serve clients standalone rather than over attached stdio.
Shipped: a signed container image (ghcr.io, SBOM + provenance + Cosign) and a Helm chart with a Restricted pod, PVC, NetworkPolicy, and PDB.
Longer term #
- More chaos classes: node pressure, network partitions, noisy neighbors, to broaden what the eval harness covers.
- Placement-aware fleet actions: proposals that target a scored subset of clusters, not one at a time.
- Industry policy presets: ready-made guardrail packs (finance, healthcare) layered on top of the base set.
How this list is chosen #
Priorities follow evidence. The eval harness surfaces where agents actually fail; those failures decide what safety work matters next. If you hit something the current design cannot express safely, open a feature request with the scenario and it becomes a candidate.
Next: Contributing.
Source: wiki/Roadmap.md — this page is generated from it, so edit the markdown, not the site.