diff --git a/src/pages/docs/release-notes.mdx b/src/pages/docs/release-notes.mdx index ca9500c1e..f04270dbe 100644 --- a/src/pages/docs/release-notes.mdx +++ b/src/pages/docs/release-notes.mdx @@ -5,6 +5,46 @@ description: "Latest Future AGI release notes covering new features, improvement {/* release-notes:insert-below — automation inserts new releases here; do not remove */} +## Week of 2026-11-02 + +
+ +
Features
+ +- **Epoch, Repair Your Agent From Failed Runs:** From the runs that failed in a simulation environment, choose **Improve** in the Debug Failures drawer and Epoch repairs the agent for you. Pick the failing cases to fix and the passing cases the fix must not break, the parts of the agent it may change, files to leave untouched, and a budget and time limit. Each candidate fix is tested as a new agent version against your cases, and the best one is then verified on a held-back set of cases it never saw while searching, so every fix ends with a verdict and an explanation of how it was decided. From there you can make the winning version your default, open a pull request for an agent connected from GitHub, or download the patch. Fixes are listed on the environment's new **Improvements** tab, where a running fix shows its steps and trials live and can be stopped at any time. Your default agent version never changes unless you choose to ship a fix. Epoch works with agents uploaded as an archive or connected from GitHub, and needs a run with enough failing cases to set some aside for verification. + +
Bugs/Improvements
+ +- **Environment Workspace Fixes:** Hover tooltips were removed from the environment workspace tabs, and an environment's scenario count now updates live instead of freezing at the first value it loaded. The environment builder chat also lets you approve all the changes in a request with one answer, while deletions and structural changes still ask for their own approval. + +- **Faster Environment Runs, More Reliable Builds:** Chat runs for agents that read their assigned port now run in parallel, cutting a 24-scenario suite from 31 to 38 minutes to 6.8 minutes. Agents that store their data in an embedded SQLite database are now supported, and environments created on an earlier platform version keep working after an upgrade. Builds and checks also no longer fail more often when many environments run at the same time. + +- **Clearer Debug Failures Results:** Debug Failures investigations now combine matching findings into a single entry, come back to findings that were set aside, and clearly label a goal that has no findings. + +- **Open the Source Trace From Eval Task Usage:** From any run on an eval task's Usage tab, Open trace now opens the exact trace with the evaluated span selected in a drawer, without leaving Usage, and previous/next steps through the task's traces in order. A voice project opens the voice call drawer instead. + +- **Error Feed Issues Reflect Only Live Eval Results:** An Error Feed issue no longer keeps listing a trace after its evaluation result has been deleted, and an issue's score, turns, and Trends statistics now reflect only live results instead of reading zero when nothing remains to measure. + +- **Eval Task List Shows Accurate Last-Run Times:** An eval task in the Tasks list now shows when it last actually produced work, instead of continuing to show its creation date no matter how many times it had run since. + +- **More Reliable Few-Shot Evaluation Inputs:** Few-shot evaluation inputs that are zero or false are no longer dropped as if they were empty, and structured (JSON) inputs are now matched and sent to the embedding model correctly instead of failing outright. + +- **Prevent Silent Blank Dataset Rows:** Writing dataset rows through the REST API, the platform's AI assistant or its MCP tools with an unrecognized column, or in a malformed batch, is now rejected with a clear error instead of silently saved as blank rows. + +- **Gateway Pricing and Model Access:** An additional public pricing source now fills gaps in the model catalogue, the model list shows only the models an API key can use, and Gemini tool calls each get their own id. + +- **Guardrail Testing and Self-Hosted Eval:** The Guardrails Test tab shows a warning for rules in Warn or Log mode, along with the real HTTP status code, and a CrowdStrike guardrail whose authentication fails is treated as a failure. On self-hosted installs, the Future AGI Eval guardrail uses your own deployment and credentials. + +- **Guardrail and Gateway Settings Stay Scoped and Intact:** Guardrail credentials, requests, webhook events, and usage are now always kept within their own organization, including when one batch spans several organizations. Saving a Guardrails or Fallbacks change keeps your other settings, such as budgets, usage limits, and IP restrictions, and guardrails can be toggled and edited for organizations managed through the Settings editor. The UI also tells you when a save or rollback did not reach the gateway. + +- **SQS Export Reliability:** The SQS export integration now retries individual failed messages and only advances past a batch once every record in it is confirmed delivered, instead of silently skipping undelivered records. + +- **Annotation Labels Filter Order:** The Labels filter toolbar now matches the Queues toolbar's order (search, then Active/Archived, then type), for consistent navigation between the two. + +- **Self-Hosted Windows Startup Fix:** Self-hosted deployments checked out on Windows now start correctly; a line-ending mismatch in shell scripts and container build files previously stopped the stack from starting. + +
+ ## Week of 2026-10-19