This Week's Highlights — issue 004, September 11, 2026

The cursor agent worker start CLI subcommand registers a machine as a self-hosted worker, opening a long-lived outbound HTTPS connection to the Cursor cloud so tool execution, codebase, build outputs, and secrets stay entirely within your own network.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ Workers are organized via 'My Machines' (single laptop/VM) or Pools/Team Pools (named queues of workers) that auto-scale and hibernate idle machines, and can run on pre-existing sandbox infrastructure including AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel, and E2B, or on custom hardware such as GPUs, Apple silicon Macs (for iOS/macOS builds), and Kubernetes. Self-hosted workers on Linux and Mac support computer use — click, type, screenshot, and browser control via Chrome or Chromium — with live desktop viewing and takeover from Cursor.

$ cursor agent worker start

Custom Modes let you pin any skill as an always-on mode in chat, activated via ⌥⏎ (Mac) or Alt+Enter (Windows) from the skill picker.

The /goal command gives a cloud agent a long-lived objective that it pursues until fully complete; pair it with a custom mode or /loop for recurring check-ins.

$ /goal fix all flaky tests and make CI green

Origin is a built-in code hosting service with repos, pull requests, code browsing, and GitHub sync, available in early beta on all paid plans. The Codebase tab is the home for Origin repos, with a +New button to create repos whose names become part of every repo URL at cursor.com/codebase/<name>; per-repo Settings manage GitHub sync status, access control, and connected apps. Cloud Agents can also start without a connected GitHub or third-party SCM provider via 'Start from scratch' in the repo picker, saving work to a Cursor Origin repo.

Projects, accessible from the left-hand nav, runs a persistent coordinator agent that plans and delegates work across thousands of parallel subagents for large bodies of work spanning months, without requiring the user to stay online. Projects run cloud-by-default on persistent compute so work continues after closing your laptop, spinning up local agents on demand for machine-specific tasks; each subagent runs on its own isolated virtual machine with a clean copy of the project and independent cloud environment. A shared context file set syncs across all cloud and local agents in a Project, growing automatically as agents record codebase knowledge and working preferences.

Subscriptions let a Project or cloud agent coordinator watch a Slack channel, run on a schedule, or follow all your PRs, acting autonomously on detected signals such as CI failures, opens, and merges without manual prompting.

Send a message to redirect an agent mid-run without interrupting it: follow-ups queue at the next tool call rather than cutting the agent off, or send immediately with 'Send now' or by pressing ⏎ twice.

Builds are pre-prepared environment snapshots (repos cloned, dependencies installed) that make cloud agents start 3x faster, with a Builds tab in the Cloud Agents dashboard for inspecting build status, logs, and commit SHAs.

Origin repos gain a per-repo Apps tab for adding app extensions — Vercel, Depot, and Buildkite — enabling preview deployments and CI integration on every PR.

Adds full CRUD for service accounts: POST /service-accounts to create one with optional initial role bindings (client secret returned only once), GET /service-accounts/{id} and GET /service-accounts to retrieve or list them, PATCH /service-accounts/{id} to update a service account's name, and DELETE /service-accounts/{id} to delete one and revoke all tokens it minted within seconds.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ A rotate-secret endpoint rotates a service account's OAuth client secret, revoking the previous secret and its tokens within seconds and returning the new secret only once.

$ curl -X POST 'https://api.pinecone.io/organizations/<org-id>/service-accounts/<service-account-id>/rotate-secret' \
  -H 'Api-Key: <api-key>'

$ curl -X POST https://api.pinecone.io/service-accounts/<service-account-id>/rotate-secret \
  -H 'Authorization: Bearer <your-token>'

Full-text search reaches general availability via the Documents API, running BM25 keyword ranking inside the same Pinecone Database index as dense and sparse vectors with no separate search cluster required. It supports Lucene query syntax (boolean operators, phrase queries, and fuzzy matching with the ~ operator) across multiple text fields per index, text-match filters that restrict semantic vector search results to documents matching a literal keyword filter, and tokenization/stemming in 18 languages plus language-agnostic n-gram tokenization for substring- and prefix-matching. Full-text search indexes support Dedicated Read Nodes for provisioned read capacity and Bring Your Own Cloud (BYOC) deployment, and are metered in the same read units and write units as vector indexes with no cluster sizing, sharding, or upgrade maintenance.

Adds a List documents endpoint that returns up to 100 documents per page in a namespace, with an optional prefix parameter to filter by ID prefix and a pagination token for iterating through results.

$ curl -X GET 'https://api.pinecone.io/namespaces/<namespace>/documents?prefix=user-&pagination=<token>' \
  -H 'Api-Key: <api-key>'

Adds lifecycle management endpoints across three resource types: organizations (List organizations, Get organization details, Update an organization name, Delete an organization), projects (List projects, Create a new project, Get project details, Update a project including maximum Pod count and customer-managed encryption key (CMEK), Delete a project), and API keys (List API keys, Delete an API key).

Adds POST /role-bindings to grant a role to a principal (such as a service account) at organization or project scope, letting a role be assigned immediately after creating the principal.

$ curl -X POST 'https://api.pinecone.io/organizations/<org-id>/role-bindings' \
  -H 'Api-Key: <api-key>' \
  -H 'Content-Type: application/json' \
  -d '{"principal_id": "<service-account-id>", "role": "editor", "scope": "project", "project_id": "<project-id>"}'

Adds auto to Claude Managed Agents permission policies, letting the server evaluate each agent or MCP tool call and automatically run it, deny it, or pause for approval; agent.tool_use and agent.mcp_tool_use events gain evaluation and evaluated_permission fields reporting how each call was evaluated under the policy.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ The ant CLI adds ant beta:sessions connect to attach your terminal to a live session — following it, sending messages, and allowing or denying pending tool calls — plus a --web flag that serves the Claude Console session viewer locally and opens the session in a browser instead.

Makes the Admin API available in the ant CLI and Python, TypeScript, C#, Go, Java, PHP, and Ruby SDKs under client.beta.organization, covering organization info, members, invites, workspaces, API keys, rate limits, service accounts, workload identity federation issuers and rules, and customer-managed encryption keys; reads credentials from ANTHROPIC_API_KEY or ANTHROPIC_AUTH_TOKEN.

Adds per-message effort control in beta on Claude Fable 5.1, Claude Mythos 5.1, and Claude Opus 5: include output_config.effort in a role: "system" message inside messages with the mid-conversation-output-config-2026-07-01 beta header to change effort mid-conversation while preserving the prompt cache. Extended to Google Cloud for the same three models.

Introduces the thinking-binding-controls-2026-08-01 beta header, which reports dropped thinking blocks in an input_transformations response field and adds thinking.block_binding.prefix_mismatch_behavior to choose between rejecting or dropping blocks whose conversation history changed.

Adds turn-scoped system messages in beta via the mid-conversation-system-clear-at-2026-08-21 header: setting clear_at: "next_user_message" on a mid-conversation role: "system" message makes it apply only to the current turn while remaining in history at no token cost.

Extends the Compliance API local session endpoints (beta, Claude Enterprise) to return transcripts for Claude Science sessions (product_surface value claude_science) and Claude for Microsoft 365 sessions in Excel, PowerPoint, Word, and Outlook (product_surface values beginning with office_agents), using the existing Compliance Access Key and read:compliance_user_data scope; also promotes the Compliance API session endpoints for Cowork and Claude Code sessions out of beta.

In Python SDK 1.2.0, TypeScript SDK 0.122.0, Go SDK 1.68.0, Java SDK 2.59.0, Ruby SDK 1.67.0, and C# SDK 12.44.0, client.beta.files and client.beta.skills no longer send the files-api-2025-04-14 and skills-2025-10-02 beta headers and return the same response shapes as client.files and client.skills; client.beta.skills.delete() now deletes a Skill together with all of its versions, and the beta Messages type BetaSkill (the container Skill reference) is renamed BetaContainerSkill.

Adds the ant apply subcommand to the ant CLI (v1.30.0) for declarative, file-based creation and updating of agents, environments, skills, memory stores, and deployments from repository files; each run writes a claude-lock.json lockfile so subsequent runs — locally or in CI — update existing resources instead of creating duplicates.

Launches claude-fable-5-1 with a 1M token context window, 128k max output tokens, always-on adaptive thinking, priced at $10/$50 per MTok input/output with prompt cache reads at $0.25 per MTok (0.025x base input price). claude-mythos-5-1 launches for Project Glasswing participants with the same context window, max output, adaptive thinking, and pricing.

Adds the "updates" value for thinking.display in beta (thinking-display-updates-2026-08-18 header), returning short progress updates Claude writes between tool calls as text, with at most one thinking block before each tool call.

Makes the computer_toolset_20260801 and browser_toolset_20260801 computer use and browser use toolsets available on Google Cloud for Claude Fable 5, Claude Mythos 5, Claude Opus 5, Claude Sonnet 5, and Claude Opus 4.8.

For accounts created on or after August 31, 2026, replaying a thinking block on Claude Fable 5.1 after the system prompt, tools, or an earlier message has changed returns a 400 error; thinking blocks produced by Claude Fable 5.1 and Claude Mythos 5.1 are also dropped by the API if replayed to an earlier model.

On Claude Fable 5.1 and Claude Mythos 5.1, tool_choice types any and tool are not supported and return a 400 error; only auto and none are accepted.

Adds concurrent eval case execution to the Java SDK (v0.3.23): Eval.start() returns an EvalResult immediately while Eval.run() still awaits completion, with lifecycle APIs getStatus(), isDone(), awaitCompletion(), getStartedAt(), getDuration(), getCasesExecuted(), and getAbortCause().⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ Concurrency is controlled via .maxConcurrency(int) on Eval.Builder or Devserver.Builder, or a custom Executor via .executor(Executor) (default: 3 concurrent cases), with global defaults settable via BRAINTRUST_DEFAULT_MAX_CONCURRENCY, BRAINTRUST_OTEL_MAX_QUEUE_SIZE, BRAINTRUST_OTEL_MAX_EXPORT_BATCH_SIZE, and BRAINTRUST_OTEL_EXPORT_INTERVAL_MILLIS. As a result, Task, Scorer, and Classifier implementations must now be thread-safe.

$ export BRAINTRUST_DEFAULT_MAX_CONCURRENCY=10
export BRAINTRUST_OTEL_MAX_QUEUE_SIZE=4096
export BRAINTRUST_OTEL_MAX_EXPORT_BATCH_SIZE=512
export BRAINTRUST_OTEL_EXPORT_INTERVAL_MILLIS=2000

Loop now runs in a Braintrust-managed runtime (server-side, not browser), persisting threads and operating across logs, experiments, and datasets project-wide; it can create and edit prompts, scorers, datasets, facets, custom views, dashboards, and automations, with each write paused for approval unless auto-accept is enabled. It runs on built-in GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna models (no AI provider config required), with GPT-6 Astra later added to the model picker including scheduled automations. An org_name parameter was added to Loop and the MCP server to disambiguate name lookups across organizations, resolving members, API keys, service tokens, patterns, prompts, and scorers by name instead of raw IDs.

Adds OpenAI Batch API instrumentation in the TypeScript SDK (v3.30.0) via openaiFilesCreateTraced, openaiBatchesRetrieveTraced, and completeOpenAIBatchTrace, tracing asynchronous OpenAI Batch requests end-to-end from file upload through submission and retrieval.

import { openaiFilesCreateTraced, openaiBatchesRetrieveTraced, completeOpenAIBatchTrace } from 'braintrust';

const file = await openaiFilesCreateTraced(openai, { file: fs.createReadStream('requests.jsonl'), purpose: 'batch' });
const batch = await openai.batches.create({ input_file_id: file.id, endpoint: '/v1/chat/completions', completion_window: '24h' });
const result = await openaiBatchesRetrieveTraced(openai, batch.id);
await completeOpenAIBatchTrace(result);

import { openaiFilesCreateTraced, openaiBatchesRetrieveTraced, completeOpenAIBatchTrace } from 'braintrust';

const file = await openaiFilesCreateTraced(openai, {
  file: fs.createReadStream('requests.jsonl'),
  purpose: 'batch',
});
const batch = await openaiBatchesRetrieveTraced(openai, file.id);
await completeOpenAIBatchTrace(batch);

import { openaiFilesCreateTraced, openaiBatchesRetrieveTraced, completeOpenAIBatchTrace } from 'braintrust';

const file = await openaiFilesCreateTraced(openai, {
  file: fs.createReadStream('requests.jsonl'),
  purpose: 'batch'
});
const batch = await openai.batches.create({
  input_file_id: file.id,
  endpoint: '/v1/chat/completions',
  completion_window: '24h'
});
await openaiBatchesRetrieveTraced(openai, batch.id);
await completeOpenAIBatchTrace(openai, batch.id);

import { openaiFilesCreateTraced, openaiBatchesRetrieveTraced, completeOpenAIBatchTrace } from 'braintrust';

const file = await openaiFilesCreateTraced(openai, { file: fs.createReadStream('requests.jsonl'), purpose: 'batch' });
const batch = await openai.batches.create({ input_file_id: file.id, endpoint: '/v1/chat/completions', completion_window: '24h' });
const result = await openaiBatchesRetrieveTraced(openai, batch.id);
await completeOpenAIBatchTrace(result);

import { openaiFilesCreateTraced, openaiBatchesRetrieveTraced, completeOpenAIBatchTrace } from 'braintrust';

const file = await openaiFilesCreateTraced(openai, { file: fs.createReadStream('requests.jsonl'), purpose: 'batch' });
const batch = await openai.batches.create({ input_file_id: file.id, endpoint: '/v1/chat/completions', completion_window: '24h' });
const result = await openaiBatchesRetrieveTraced(openai, batch.id);
await completeOpenAIBatchTrace(result);

Adds Agno eval instrumentation to the Python SDK (v0.36.0) — AccuracyEval, AgentAsJudgeEval, ReliabilityEval, PerformanceEval, and eval suites are traced automatically via auto_instrument(); eval suite runs open a Braintrust experiment automatically.

import braintrust
from agno.eval.accuracy import AccuracyEval

braintrust.auto_instrument()

eval_run = AccuracyEval(agent=my_agent, dataset=my_dataset)
eval_run.run()

from braintrust import auto_instrument
from agno.evals import AccuracyEval, AgentAsJudgeEval, ReliabilityEval, PerformanceEval

auto_instrument()

eval_suite = [
    AccuracyEval(agent=my_agent, questions=questions),
    AgentAsJudgeEval(agent=my_agent, rubric=rubric),
    ReliabilityEval(agent=my_agent, scenarios=scenarios),
    PerformanceEval(agent=my_agent, tasks=tasks),
]
for e in eval_suite:
    e.run()

from braintrust.integrations.agno import auto_instrument
from agno.evals import AccuracyEval, ReliabilityEval

auto_instrument()

AccuracyEval(agent=my_agent, inputs=test_inputs).run()
ReliabilityEval(agent=my_agent, inputs=test_inputs).run()

Adds has_error, metadata, and duration fields to SpanFilters for trace.get_spans() in the Python SDK (v0.39.0), letting trace-level scorers filter spans by error state, metadata values, and execution time bounds.

spans = trace.get_spans(SpanFilters(has_error=True, duration={"gt": 5.0}))

Adds report_scores and report_metrics inputs to the GitHub eval action (v2.1.0) to filter which scores and metrics appear in the PR comment, each accepting a comma- or newline-separated list of names.

- uses: braintrustdata/[email protected]
  with:
    api_key: ${{ secrets.BRAINTRUST_API_KEY }}
    report_scores: accuracy, pass_rate
    report_metrics: latency, cost

- uses: braintrustdata/eval-action@v2
  with:
    report_scores: accuracy,faithfulness
    report_metrics: latency,cost

- uses: braintrustdata/eval-action@v2
  with:
    api_key: ${{ secrets.BRAINTRUST_API_KEY }}
    report_scores: accuracy, relevance
    report_metrics: latency, cost

- uses: braintrust-dev/[email protected]
  with:
    api_key: ${{ secrets.BRAINTRUST_API_KEY }}
    report_scores: accuracy, faithfulness
    report_metrics: latency, cost

- uses: braintrustdata/[email protected]
  with:
    experiment_name: my-experiment
    report_scores: accuracy, coherence
    report_metrics: latency, cost

Adds ElevenLabs instrumentation (wrapElevenLabs, TS SDK v3.32.0) for tracing text-to-speech and speech-to-text calls including streaming and timestamped audio, and @google/generative-ai instrumentation (wrapGoogleGenerativeAI, TS SDK v3.32.0) for the legacy Google Generative AI SDK (v0.24.x).

import { wrapElevenLabs } from 'braintrust';
import ElevenLabs from 'elevenlabs';

const client = wrapElevenLabs(new ElevenLabs({ apiKey: process.env.ELEVENLABS_API_KEY }));

Patterns runs Loop against a project on a schedule to surface recurring problems, cost trends, and cohort-level failure modes not covered by any scorer; each finding is saved with supporting traces and a suggested fix, and can be used to create a scorer or classifier, continue investigation in Loop, or be copied as a prompt. Loop automations power Pattern discovery by default, running daily at 9:00 AM with timezone selection for custom cron schedules.

Adds glm-5.3-flash, a multimodal reasoning model served by Braintrust with no AI provider setup required, selectable under the Braintrust provider in playgrounds, prompts, and scorers, or requestable via glm-5.3-flash through the Braintrust Gateway.

$ curl https://api.braintrust.dev/v1/proxy/chat/completions \
  -H 'Authorization: Bearer <BRAINTRUST_API_KEY>' \
  -H 'Content-Type: application/json' \
  -d '{"model": "glm-5.3-flash", "messages": [{"role": "user", "content": "Summarize this trace for me."}]}'

$ curl https://api.braintrust.dev/v1/gateway/chat/completions \
  -H 'Authorization: Bearer <BRAINTRUST_API_KEY>' \
  -H 'Content-Type: application/json' \
  -d '{"model": "glm-5.3-flash", "messages": [{"role": "user", "content": "Score this response for accuracy: ..."}]}'

Loop automations are scheduled Loop runs with their own instruction, model, and write permissions that produce a read-only thread you can Continue, with results deliverable to a Slack channel or webhook; automations support timezone selection for custom cron schedules.

Adds standard verifier output upload to the Harbor plugin in the Python SDK (v0.36.0): test-stdout.txt, test-stderr.txt, and ctrf.json, with attachments and redact_patterns configuration.

Allows scorer functions in the Python SDK (v0.38.0) and TypeScript SDK (v3.31.0) to return {"score": 0.5} without a name field, using the function's name as the score key.

Adds prompt support to the Go SDK (v0.13.0): load a Braintrust prompt, render its variables locally, call any LLM client with the result, drive eval parameters, and link rendered calls back to their prompt in Braintrust traces.

The 'Blind human reviews' project setting hides peer scores, comments, and aggregates until a reviewer submits their own scores; reviewers holding the project Update permission are always exempt.

Adds named, reorderable, collapsible, duplicable dashboard sections with per-browser collapse-state memory (Pro and Enterprise plans), plus a shareable fullscreen link for individual dashboard charts, with chart move and copy actions now in each chart's more-options menu.

Introduces the vllm-tt-plugin out-of-tree platform plugin that automatically registers Tenstorrent accelerators as a vLLM platform when ttnn from TT-Metal is importable, supporting TTLlamaForCausalLM, TTMllamaForConditionalGeneration, TTQwen2ForCausalLM, TTQwen3ForCausalLM, TTQwen3_5ForConditionalGeneration, TTQwen2_5_VLForConditionalGeneration, TTQwen3VLForConditionalGeneration, TTMistralForCausalLM, TTMistral3ForConditionalGeneration, TTGemma3ForConditionalGeneration, TTGemma4ForCausalLM, TTGemma4ForConditionalGeneration, TTGemma4UnifiedForConditionalGeneration, TTDeepseekV3ForCausalLM, and TTGptOssForCausalLM.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ Tenstorrent mesh targets are selected via the MESH_DEVICE setting (e.g. TG), which replaces --tensor-parallel-size / -tp / -pp (rejected outright on TT targets); --additional-config.tt.sample_on_device_mode (e.g. all) enables on-device sampling so the host receives already-chosen tokens rather than raw logits, and --additional-config.tt.fabric_config configures the mesh fabric topology (FABRIC_1D, FABRIC_2D, FABRIC_1D_RING). Scheduling uses a phase-constrained TTScheduler that issues exclusively prefill-only or decode-only steps (with chunked prefill and interleaved decode), and an alternative TTLaneCoordinator scheduler class (scheduler_config.scheduler_cls) runs one independent TTScheduler per data-parallel lane for Galaxy-class single-execute models, with TT_LLAMA_TEXT_VER (e.g. llama3_70b_galaxy) and TT_QWEN3_TEXT_VER (e.g. qwen3_32b_galaxy) selecting Galaxy-optimised single-execute model variants.

$ vllm serve meta-llama/Llama-3.3-70B-Instruct --additional-config.tt.sample_on_device_mode all --additional-config.tt.fabric_config FABRIC_1D_RING

$ TT_QWEN3_TEXT_VER=qwen3_32b_galaxy vllm serve Qwen/Qwen3-32B --additional-config.tt.fabric_config FABRIC_2D

Adds EXTRA_MODELS_DIR environment variable to register additional model architectures at startup from bundle folders each containing a vllm_metadata.json and an adapter class, without editing plugin source, and TT_VLLM_BUILTIN_MODELS=0 to restrict the architecture registry to only models supplied via EXTRA_MODELS_DIR, excluding all built-in models.

$ EXTRA_MODELS_DIR=/opt/tt-models TT_VLLM_BUILTIN_MODELS=0 vllm serve /opt/tt-models/my-custom-model

Adds --max-num-queued-reqs and --max-num-queued-tokens CLI flags for admission control on the request queue, capping queued load to prevent runaway queue growth during traffic spikes.

$ vllm serve meta-llama/Llama-3.1-8B-Instruct --max-num-queued-reqs 500 --max-num-queued-tokens 131072

Adds support for new models: Hy4-preview (Tencent 770B/49B-active MoE with Gated DeepSeek Sparse Attention and native MTP), Qwen3.8-Flash-Next (BF16/FP8/NVFP4 with MTP), GraniteSWA and GraniteMoeSWA, NemotronH_Omni_Reasoning_V3 with MTP, and Kimi K3 NVFP4 checkpoints.

Adds sharded_rdt P2P backend for RL weight sync, where each worker pulls only its TP/EP slice over NIXL or Ray Direct Transport.

Adds prefix_cache_retention_interval as a CLI argument (default 0) to control Mamba prefix cache retention, makes prefix-cache NONE_HASH deterministic by default (removing the need to pin PYTHONHASHSEED for distributed KV cache users), and introduces a KVCacheLayout enum to standardize KV cache layout across backends.

Adds --linear-backend flashinfer_cutedsl opt-in flag to enable FlashInfer CuTeDSL BF16 low-latency GEMM, and enables FlashInfer all-reduce by default for tensor-parallel CUDA groups, with VLLM_ALLREDUCE_USE_FLASHINFER=0 to opt out.

Removes the PyAV video decoder backend, deprecates python -m vllm.entrypoints.openai.api_server in favor of vllm serve, and removes the VLLM_TEST_FORCE_FP8_MARLIN environment variable.