1. Docs
  2. Changes and releases
  3. AI Agents

AI Agents

  • AI Agents agent-platform v1.6.1

    Fixed

    • Select agentgateway/muster by stable labels in connectivity netpols in #156 by @teemow

    Full Changelog: https://github.com/giantswarm/agentic-platform/compare/v1.6.0...v1.6.1

  • AI Agents agent-platform v1.6.0

    Added

    • Remediation retries on rendered child HelmReleases in #155 by @teemow

    Full Changelog: https://github.com/giantswarm/agentic-platform/compare/v1.5.1...v1.6.0

  • AI Agents agent-platform v1.5.1

    Fixed

    • Make meta-package reconcile on flux-multi-tenancy clusters in #154 by @teemow

    Full Changelog: https://github.com/giantswarm/agentic-platform/compare/v1.5.0...v1.5.1

  • AI Agents agent-platform v1.5.0

    Added

    Full Changelog: https://github.com/giantswarm/agentic-platform/compare/v1.4.13...v1.5.0

  • AI Agents muster v0.10.0

    Added

    • (workflow) Decouple referencing from output, add output projection, unify expression language in #876 by @teemow

    Full Changelog: https://github.com/giantswarm/muster/compare/v0.9.1...v0.10.0

  • AI Agents muster v0.9.1

    Fixed

    • (metatools) Lower discovery default filter limit to 5 in #878 by @teemow

    Full Changelog: https://github.com/giantswarm/muster/compare/v0.9.0...v0.9.1

  • AI Agents agent-platform v1.4.13

    Changed

    • Bump muster sub-chart 0.8.4 -> 0.9.0, which adds a cheap, ranked, faceted tool-discovery tier (giantswarm/muster#868): filter_tools is now a discovery tier distinct from execution, so finding a tool no longer scales with the full descriptive weight of every candidate. It gains limit/offset (default page 25) returning total + truncated, defaults to a one-line summary per tool with no input schema (full description/schema stay available via describe_tool or include_schema=true), adds a BM25-ranked query mode that returns matches best-first with a score, and exposes Workflow CRD metadata.labels as a filterable labels facet. Against a ~280-workflow fleet a broad filter_tools(pattern="*workflow*") call drops from a ~330 KB full-catalogue dump to a ~3 KB summary page (~100x smaller). The companion muster-crds / agentic-platform-crds chart is intentionally NOT bumped: #868 propagates existing Kubernetes labels onto tool metadata and changes no CRD schema, so muster 0.9.0 runs against the existing 0.8.4 CRDs.
  • AI Agents muster v0.9.0

    Added

    • Cheap, ranked, faceted tool discovery tier (#868): filter_tools is now a discovery tier distinct from execution, so finding a tool no longer scales with the full descriptive weight of every candidate. Against a ~280-workflow fleet a broad filter_tools(pattern="*workflow*") call returns a bounded summary page (~3 KB) instead of the full-catalogue dump (~330 KB) — measured ~100x smaller.
      • Bounded + summarised pages. filter_tools gains limit (default 25) and offset, and the response carries total and a truncated flag. Discovery now defaults to a one-line summary per tool with no input schema; the authoritative full description and schema remain available via describe_tool, or by passing include_schema=true.
      • Ranked query mode. A new query argument relevance-ranks matches with a dependency-free lexical ranker (Okapi BM25 over name + summary) and returns them best-first with a score, dropping non-matching tools. Lexical ranking needs no embedding index; embeddings can be a later upgrade.
      • Label facets. Workflow CRD metadata.labels are now propagated onto the workflow’s execution tool and can be filtered in discovery via a labels facet (key=value; all must match), letting clients scope a lookup to a labelled subset.
      • Agent REPL parity. The filter command (aliases find/search) now exposes the discovery tier: options are given as key=value pairs (pattern, description, query, labels as k=v,k2=v2, case_sensitive, detailed, limit, offset), with a bare token still accepted as the name pattern. It now reports the page size against the total match count and the catalogue size, and prints the next offset when more matches exist (previously it mislabelled the page size as the match count and hid truncation). Unknown options and stray extra arguments now fail loudly with the list of valid keys instead of being silently ignored, the name column auto-sizes to the longest tool name on the page, and tab-completion only offers options that have not been supplied yet.
      • Existing list_tools / call_tool / describe_tool behaviour is unchanged, as is list_core_tools (which keeps full descriptions and schemas).
  • AI Agents agent-platform v1.4.12

    Changed

    • Bump muster (and muster-crds) sub-chart 0.7.5 -> 0.8.4, which adds workflow control flow (giantswarm/muster#865): forEach (sequential loops over a list), parallel (concurrent sub-step groups with isolated contexts), workflow-level onFailure (best-effort cleanup/rollback), and condition.template (boolean Go-template step gates). The muster-crds chart carries the matching Workflow CRD schema (new spec.steps[].forEach/parallel, spec.onFailure, condition.template; the dead step.outputs field is removed in favour of store: true + {{ .results.<step_id> }}), so the CRDs chart must be upgraded alongside the app chart. Also fixes camelCase field-name parsing (allowFailure/fromStep/expectNot/jsonPath) in the structured workflow_create/update/validate tool path. The 0.8.4 bump (over 0.8.0) carries giantswarm/muster#871: the 0.8.0 Workflow CRD was rejected by the Kubernetes API server on k8s >= 1.34 because its “exactly one of” CEL guards used a [...].filter(x, x).size() == 1 form whose estimated cost exceeded the schema-wide validation budget (3.2x over), which made agentic-platform-crds fail to install/upgrade and roll back. Rewriting the guards to the cheap additive form fixes the install.
  • AI Agents agent-platform v1.4.11

    Changed

    Full Changelog: https://github.com/giantswarm/agentic-platform/compare/v1.4.10...v1.4.11