Changelog
C3 AI Agentix changelog
This page collects the user-facing changes for each C3 AI Agentix release since 1.0.2, newest first. See Release Notes for an overview of what shipped at general availability.
Internal changes, such as tests, refactors, and dependency updates, are left out of every version below, as are changes to Document AI, which releases on its own schedule with its own changelog. Features still behind a flag are opt-in and not yet released, so they're not listed here until they reach general availability.
1.0.5
Fixed
- Fixed agent memory: the agent now remembers your role, your naming conventions, the preferences you stated once, and the decisions you already made, so you stop re-explaining your context in every new chat. Memory is scoped to you, and you manage it in conversation by telling the agent to remember, correct, or forget something. If the memory store is unavailable, chats still start, just without your history.
- Fixed automations: you can now create and manage scheduled automations by talking to the agent, and an automation run falls back to a working model route when its configured model has none.
- Fixed datasets: you can publish a derived dataset to the Dataset Library, delete it durably, and rely on a reconciliation sweep to repair drift. Hosted agent backends other than OpenCode can fetch attached datasets, and resuming a chat no longer skips chats whose dataset grants were still being replayed.
- Fixed chat history and traces: the chat turn list, turn summaries, and conversation and query history now read from Laminar with paging, per-turn generated artifacts render inline in the conversation, retrieved resources appear in each turn summary, and LLM and tool trace spans are ordered on one clock.
- Fixed LLM guardrails and safety: you can apply your own Bedrock Guardrail to LLM proxy traffic, administrators get controls for it, and a guardrail failure no longer takes down unrelated requests. Optional safety-review routing sends sensitive requests through a review model with its own usage accounting.
- Fixed LLM routing: stalled Bedrock streams time out instead of hanging, transient proxy transport failures are retried during route revalidation, LLM route provisioning retries promptly when the proxy isn't ready, and an inconclusive status check no longer marks a healthy route stale. Router failures keep a sanitized reason that shows in status views, and router usage by user can be sorted and exported to CSV.
- Fixed model configuration: models that report a fractional input-token limit are accepted, Foundry GPT routes are priced correctly, Azure AI routes support a custom API base and can be re-pointed in place, the model catalog snapshot fits its storage, and reasoning is preserved across the Messages and Responses APIs.
- Fixed agent session lifecycle: resuming sessions are listed above stopped ones in the dashboard inventory, an idle stop no longer leaves a volume behind, stale approval and tool prompts no longer hide later prompts, a stale snapshot can't overwrite live chats on resume, a lost instruction bundle is recovered on demand for remote-hosted agents, and Foundry chats keep their workspace files across idle compute deprovisioning.
- Fixed agent capacity and cleanup: concurrent pool-pod boots are capped for waiting reserves, a reserve fails fast on repeated pool errors, automation run pods save a snapshot and are reclaimed on stop, like chats, terminated interactive agent services are deleted after 15 days, and agent pods drain before a self-upgrade restarts the app.
- Fixed chat and workspace: the workspace panel explains why it's empty on a reloaded chat and retries the file tree while a service resumes, chat users can call web search, a malformed question message renders nothing instead of an error, and the spend and cap column on the Access users page refreshes after budget edits.
- Fixed admin and settings: Admin settings are regrouped into one rail with a new Data & privacy page, the History tab is removed from the Agent App, and expected authorization failures in Studio log as a single line instead of a stack trace.
- Fixed agent skills: the skill catalog fits the harness listing through a bounded index, skills no longer treat app-listing commands as shell commands, search and RAG queries are produced in the language of your data, and a new skill helps you talk to the agent while you author one.
- Fixed Studio package access, so a package zip's manifest reads correctly, and fixed app identity provider setup through new application administrator APIs.
1.0.4
Fixed
- Fixed custom agent publication: an agent published from another app now shows up in your agent list as a read-only Downloaded snapshot, tagged with the app it came from. Open it and chat with it straight away, or clone it into a regular, fully editable agent of your own. The clone deep-copies the snapshot's skills and plugins into editable copies, so you can tune it without touching the original. Downloaded snapshots stay out of the skill and plugin libraries, and publishing creates the reader role when it's missing.
- Fixed MCP Apps: agent-generated HTML artifacts can now call back to the app while you use them, so a rendered dashboard can query live data, read sibling workspace files, write data back, or send a follow-up message to the agent, all through an isolated bridge that never gives the artifact direct access to your session. Third-party MCP App servers, such as a hosted whiteboard app, can also run their tools and render their UI directly in chat.
- Fixed agent plugins: bundled plugins now count as visible in agent plugin assignment, and base users can open the Plugins page.
- Fixed automations: you can cancel a running job run, see the last run status on the Automations list, follow the trace link and live usage on the run detail page, and jump to all runs from the Recent runs card. Enabled jobs are listed first, and an unattended run that stalls reports the question it was waiting on.
- Fixed agent services stuck or crash-looping at startup: stuck and crash-looping sessions are detected and terminated earlier, non-terminal startup faults surface as an early signal, hibernated pods are recovered at the automatic resume cap, and terminated chats recover on send instead of dead-ending.
- Fixed long-chat continuity: compaction and catalog reconciliation are hardened, a chat's transcript is verified as restored before it's moved, and model catalog reconciliation is faster.
- Fixed upgrades: rolling updates stop first with an automatic credential upgrade, a proxy worker with a snapshot no longer wedges the credential upgrade, the Upgrades panel no longer flickers after scheduling, the service health banner stays out of the way during an upgrade, cloud service operations are rejected while a self-upgrade is in progress, and the "upgrade not marked complete" banner shows an error ID.
- Fixed a false post-upgrade verification failure on healthy upgrades.
- Fixed LiteLLM and model routing: prompt caching is enabled on personal API keys, remote-hosted agents get an LLM route through the proxy sidecar, routing for Claude Code, Cursor, and other coding tools holds across non-Anthropic providers, app admins can set Bedrock keys, and Bedrock routes handle stop sequences and preserve native thinking blocks.
- Fixed agent administration: administrators can promote skills to Global, personal skills show a hint that they can be added from the skill's page, the Admin tab is restricted to application administrators, and the admin service log panel strips terminal escape codes.
- Fixed MCP access for confidential clients by adding OAuth client secrets.
- Fixed admin and chat surfaces: less polling, no flicker on the Agent Services dashboard, no stray scroll bar under the admin page, and banner hide and dismiss controls explain that hidden banners can be restored. Role lists show a tooltip when they exceed the visible count.
- Fixed the UI branding: several locale files now say Agentix.
- Fixed downloaded chat transcripts so they begin with the agent's welcome message, and the Embed chat panel now shows how the npm-registry delivery path works.
- Fixed service details showing "starting" for stopped services.
- Fixed chat access when a service owner was cleared.
- Fixed a routing activation timeout for job sessions.
- Fixed MCP app window icons overlapping the side panel, and the display mode for inline MCP apps.
- Fixed additional-auth model sync discovery and admin panel behavior, and model selection dropdown scrolling in the API access setup modal.
- Fixed the agent pod lease going stale after a tab regained focus, and the admin page returning 403 during an app upgrade.
- Fixed pod re-pointing on resume so a language change no longer cancels it permanently, and chat instructions are re-delivered after a resume on remote backends.
- Fixed a text-only turn being rejected when it carried no dataset IDs.
- Fixed the Agentix package update overwriting the server version.
- Fixed model context overflow: the context guard is armed on the canonical routes, and a request that exceeds a model's context window now returns a prompt-too-long error that clients recognize.
- Fixed model configuration: duplicate Bedrock reasoning effort fields are reconciled, default-On model intent is preserved when the catalog refreshes, and enabled provider slots sort to the top with the primary first.
- Fixed Azure model routing: gpt-5.4 and later models route through the responses upstream, and Azure deployments map to LiteLLM routes from the auth config.
- Updated the recommended C3 server version to 8.9.17+413.
1.0.3
Changed
- Agent plugins: package agent extension behavior (skills, hooks, files, prompts, and config) as a versioned, reusable plugin, upload it as a single zip, and attach it to an agent definition for everyone or to just your own sessions.
- Team spend management: administrators can group users into teams that share one LLM budget and hand day-to-day allocation to a team admin, with each member able to see their own allocation and current spend.
- Default chat models updated to the latest Claude Sonnet and Opus releases.
- Faster startup for the web application; the LLM service Models page no longer polls the server twice a second.
- The LiteLLM model catalog refreshes in a running agent pod with no restart, and a slow catalog load during startup is treated as a transient condition rather than a hard failure.
- Chats are recovered when their pod is idle-reclaimed or rotated, and warm sessions resume faster without spurious "history expired" errors.
- Resuming a chat now stops your other active session, so one conversation stays in one place.
- New REST endpoints for skills and connections extend the Agent REST API.
- A stopped session's workspace is protected from reclamation when it resumes, and a serving endpoint is never dropped in favor of an unproven replacement.
- More resilient chat delivery against transient gateway errors, and agent startup diagnostics were moved off the readiness path for faster availability.
- The Updates page now shows the server version, mode, and node layout; agent detail pages show hardware profile and resource usage.
- Personal skills can be assigned to selected agents and appear on the Agent Detail page, with stale assignments cleaned up when a skill is removed.
- Route assignments are visible to all users and scoped to the applications each user can access.
- Automation runs dispatch on the agent's configured model, and the agent, skill, plugin, and connection names are frozen onto each run when it's dispatched.
- Model catalog discovery explains why a model was skipped, and Azure catalogs publish with a usable default chat model.
- The LLM API view shows total user and team counts.
- Starting a new chat is blocked when no working models are available, instead of failing silently.
- Confirmation windows stay open and long tool-call text stays readable during a failed action; the Automations link in the New Automation Job Setup flow can be selected.
- Clearer status messaging: exception stack traces are hidden from automatic-update details, and background-action errors are surfaced with cleanup.
- Completed agent containers are classified by their role instead of being reported as errors, and stopped sessions are excluded from the session-pool failed count.
Fixed
- Fixed a duplicate message appearing in chat under a specific timing condition.
- Fixed a workspace side-panel race during cold navigation.
- Fixed an issue where deleting one asset could remove unrelated assets that shared its name prefix.
- Fixed model-selection issues and removed a redundant default entry from the model picker.
- Fixed permission errors for administrators on the admin dashboard surfaces and when listing environment applications.
- Fixed admin usage totals that could be quietly truncated; the report is now assembled from durable daily slices.
- Fixed code blocks that didn't render correctly in the skills preview.
- Schedule editing now uses local time; fixed instruction text that overflowed its container.