The wire

DSPy 3.4.1 promotes GEPA's structured InstructionProposer to the default (routing reflective examples through dspy.Predict + JSONAdapter so History, tool trajectories, and dspy.Image stay structured) and lets dspy.RLM build DSPy sub-agents (Predict, ChainOfThought, ReActV2, nested RLM) directly from its sandboxed REPL, sharing one max_llm_calls budget. Adds Python 3.15, fixes 3.4.0 import crash.

The official Model Context Protocol reference servers monorepo cuts its first stable 1.0.0 release. The TypeScript servers @modelcontextprotocol/server-{everything,filesystem,memory,sequential-thinking} ship at 1.0.0 on npm; the Python servers mcp-server-{fetch,git,time} ship at 2026.10.10 on PyPI. Includes major filesystem roots/Unicode/UNC fixes, memory graph concurrency and validation hardening, and Streamable HTTP/SSE session fixes aligned with MCP SDK 1.x.

The official Anthropic Python SDK v1.13.0 adds API types for Chat and Cowork unified analytics metrics, plus support for workflows, multiagent configuration, and thread status filtering on Managed Agents. Also names the default API version header value internally.

Claude Code 2.1.295 adds blocking hook failure mode (`onFailure: "block"`), Program Status Protocol (OSC 7501) terminal signaling, per-upstream model allowlists on the Claude apps gateway, a `models` wildcard list, `forceLoginMethod: "gateway"`, retry watchdog max wait cap, and native mod notifications. Also fixes remote MCP reconnection backoff, MCP pagination loops, and many agent/subagent handoff bugs.

Claude Code 2.1.293 ships Claude Haiku 5.5 (`claude-haiku-5-5`) as the default Haiku model on the Anthropic API, with a 1M token context and pricing of $0.10/$0.50 per Mtok ($0.50/$2.50 above 100K). The release also adds `agentType` to the subagentStatusLine payload, an `isDeferred` flag on `$.tool.register` so mods can expose a tool schema up front instead of behind tool search, and fixes a memory leak in HTTP MCP connections plus many agent/subagent and session bugs.

Anthropic's Python SDK v1.12.0 ships support for the new claude-haiku-5-5 model, adds typed tool calls for the computer and browser toolsets, exposes web_search and code_execution in model capabilities, adds url_sources to the Managed Agents web_fetch tool config, surfaces a lifecycle filter and line field on /v1/models, and documents the 256 agent-tools limit.

Anthropic announced Claude Haiku 5.5, positioned as its fastest, cheapest, and most capable small model to date. The model targets high-volume, latency- and cost-sensitive agent workloads such as classification, routing, extraction, and tool-use loops where price-per-token and throughput dominate the decision to deploy.

Claude Code 2.1.292 ships several agent-relevant additions: a new `effort` parameter on the Agent tool to control sub-agent effort level, `--marketplace <source>` on `claude plugin install` to add + install in one step, prompt caching in the `$.model.complete` mod API via `cache: true` blocks, a `prompt.autocomplete` hook, workflow agents exposed to `agent.spawn`, plus many security fixes (UNC path reads, sandbox escapes, tampered settings cache) and reliability fixes for bac

The OpenAI Python SDK 3.25.0 extends its agents surface: local tool failures are now observable, local tools can run during streamed session creation, and the API adds agent turn items plus usage source grouping. Typed function helpers also support `defer_loading=True` tool search, and `type_to_response_format_param` is now exported publicly.

Transformers 5.19.0 ships EmbeddingGemma 2, a Gemma-4-based multimodal embedding model that encodes text, images, audio, and video into a shared 768-d space with Matryoshka truncation to 512/256/128 dims and configurable vision/audio towers. The release also adds expert-parallel token dispatch (default for Qwen3 MoE/Mellum), per-layer cache configuration, continuous-batching support in regular SDPA/flash attention, and deprecates the "paged|" attention prefix.

Claude Code 2.1.290 extends plugin/mod hooks with `serverToolUses` (advisor tool calls with ids, inputs, timing), adds `agentId` to `tool.check` to distinguish subagents, adds `ceiling` for org-required approvals, and new `ThemeKey`/`Color` typings. CLI gains `claude attach <name>` and `claude logs <name>` by partial session name, plus `/claude-api managed-agents-onboard` to scaffold Managed Agents (e.g. `deep-researcher`). WebFetch now reports truncation and accepts `offset`

vLLM 0.31.0 ships 717 commits with DeepSeek-V4.1-Flash FlashMLA + NVFP4 compressed KV as SM100 default, a new `vllm preload` daemon that keeps post-quantized weights resident across engine restarts (plus CRIU-based `vllm snapshot create/restore`), Model Runner V2 draft-model speculative decoding with custom logits processors, MoonEP balanced EP all2all, prefill context parallelism, new scheduling controls, and tighter security on per-request multimodal kwargs.

The official OpenAI Node SDK v7.28.0 adds API support for creating custom voices and emits new agent session events, expanding what TypeScript/JavaScript agent builders can do with OpenAI's voice and Agents APIs.

MCP Python SDK v2.3.0 adds max_sse_event_size on Streamable HTTP clients, an MCPServer(subscriptions=False) opt-out, a get_tool_input_schema hook for low-level servers, and one-shot client retry after a HeaderMismatch (-32020). Also stops counting interactive OAuth logins against request timeouts, rejects invalid x-mcp-header annotations at registration, and omits empty _meta/params/experimental fields on legacy connections.

Claude Code 2.1.288 adds a `$.ui.selection()` API for mods to read the user's selected text/transcript row, a re-authenticate prompt when an MCP server needs more OAuth scope mid tool call, a `--max-findings <n>|all` flag for `/code-review`, built-in `gh api` in cloud sessions, Ctrl+C draft recovery, and many fixes around resume, auto-compaction, MCP double-calls over 16 MB, subagent tool inheritance, and sandboxed Bash safety.

Model Context Protocol TypeScript SDK 2.3.0 ships stateless Streamable HTTP (one server per request), same-origin redirect policy on HTTP client transports, new `maxToolInputElements` and `expectedResource` audience checks for bearer auth, support for `tasks/get`/`tasks/cancel` on 2026-07-28 connections, browser-extension origin allow-listing, and major perf wins for large SSE messages (100MB tool result went from ~60s/1.7GB to <1s/0.5GB).

OpenAI Agents SDK for Python v0.23.0 adds configurable MCP listing page limits, opt-in Docker sandbox removal protection, configurable sandbox memory consolidation turns, and an opt-in encrypted session history scan budget. The release also ships numerous core fixes around nested agent-tool state isolation, tool identity preservation, guardrail diagnostics persistence, and redaction of streaming/tool failure tracebacks.

The OpenAI Python SDK 3.23.0 ships several agent-focused features: typed answers from session streams, staging files and downloading turn artifacts, exposing typed application actions as agent tools, collecting final output from beta Agents streams, plus new session traces and Realtime translations endpoints. Also fixes response reconnect behavior and the eval run cancellation endpoint.

The OpenAI Node SDK 7.26.0 ships two agent-relevant features: listing traces for an agent session, and a beta API to bind typed application functions directly to Agents tools. Also fixes WebSocket credential refresh on reconnects and lane ID reservation across session replacements for the Responses API, plus corrects the eval run cancellation endpoint.

Claude Code 2.1.287 introduces Claude Mods (plugins can modify deeper behavior), a new opt-in "You Should Know" side-agent plugin that flags things you or Claude might miss, an `n:<text>` filter in the agents view, support for URL prompts from MCP servers on the 2025-11-25 protocol (e.g. sign-in flows), `prompt_text` added to OpenTelemetry events, plus dozens of fixes across MCP, hooks, SDK streaming, Bedrock/Vertex, and screen reader mode.

Anthropic's Python SDK v1.10.0 exposes several new agent-facing surfaces: refusal stop_reason and stop_details on Managed Agents session idle events, an MCP Tunnels beta with a read-only transport object and one-time relay token, Plugins/Plugin Marketplaces in the Admin API, per-user usage & cost analytics, Claude Enterprise RBAC groups/roles, and GA for Organization API endpoints. Also fixes session tool runner turn-ending and auto-pagination edge cases.

Transformers 5.18.0 adds four new model families: Nemotron 3 Diarization (streaming speaker diarization up to 8 speakers with configurable 80ms–30.4s latency), NemotronH Omni (Mamba-Transformer LM + RADIO vision + Parakeet audio for joint text/image/video/audio reasoning), HyperCLOVAX Vision V2 (NAVER multimodal VLM with think tokens), and GTE (RoPE BERT-style long-context multilingual embeddings backing gte-*-v1.5 and Snowflake arctic-embed-m-v2.0).

Ollama v0.35.1 adds support for Cloudflare's open-source Clef (27B) and Clef Flash (9B) decision models via a new /v1/systemone endpoint that scores multiple yes/no questions jointly against a shared text+image state. Web-search models can now perform up to 10 searches per response, Modelfiles gain a CAPABILITY declaration preserved through inheritance/export, and decision models now report only the `decision` capability.

Claude Code 2.1.285 ships several new agent controls: `claude --desktop` opens the Claude desktop app on the current dir or a resumed session, `claude plugin configure` inspects/sets plugin options (including bundled .mcpb MCP server settings at install time), a new `allowedProviders` managed setting restricts which API providers (Anthropic, Bedrock, Vertex, Foundry, etc.) a machine can use, plus `CLAUDE_CODE_DISABLE_WEB_FETCH` and retry-cap env vars. Also many fixes to subag

The official OpenAI Node SDK ships v7.24.0 with several agent-facing features: Agents API credential and session options, a new GPT-6.1 Sol model identifier, Realtime Translation client-secret creation with a dedicated translation WebSocket, Cyber access programs for Responses, and opt-in incremental output snapshots streamed over WebSocket for the Responses API.

OpenAI announced GPT-6.1 Sol, a new model positioned just below its flagship Astra tier. Sol targets coding, computer use, and professional/knowledge work while charging roughly one-fifth of Astra's standard API input and output token prices, making high-quality tool-using agents substantially cheaper to run at scale.

Ollama 0.35.0 introduces support for "decision models" through a new `/v1/systemone` API, based on TypeSafe's Jev API. Instead of free-form text, these models return structured choices, probabilities, and confidence scores — useful for ticket triage, model routing, and classification. Ships with Bespoke Labs' Nimble and Together AI's Tev1. Supports `choice`, `noul`, and `score` question types.

MCP TypeScript SDK 2.2.0 ships auto-paginated list calls (listTools/listPrompts/listResources follow nextCursor until done, capped by listMaxPages), tightens machine-to-machine OAuth by requiring expectedIssuer on ClientCredentials/PrivateKeyJwt/CrossAppAccess providers, and adds AuthorizationServerMismatchError so fetchToken() fails fast when client info is bound to a different auth server. Includes CJS type fixes and various server/handler fixes.

Anthropic's Python SDK 1.9.0 adds claude-sonnet-5-5, a new between_tools extended-thinking type, and an option to run tool calls in parallel while the assistant reply is still streaming. Cache diagnostics on Message / MessageCreateParams are now GA, workspace rate limits gain include_inherited/source, and Managed Agents event filters get typed event values. Includes fixes to stream()/parse() diagnostics parsing.

Claude Code CLI v2.1.284 makes Claude Sonnet 5.5 (`claude-sonnet-5-5`) the default Sonnet model on the Anthropic API — 1M context, $2/$10 per Mtok with $0.20/Mtok cache reads. Release also adds `/mcp reconnect all`, a "yes but ask again" auto-mode option, dollar-denominated spend limits in `/usage`, `effortSlider` keybindings, Google Cloud OTLP telemetry export, and many MCP/Agent SDK bug fixes.

Vespper (YC F24) released a hosted MCP server that lets AI agents edit Microsoft Word (.docx) files as if they were HTML. The agent reads/searches/edits an HTML projection of the doc, and Vespper's fine-tuned 3-8B LoRA reconciler model maps HTML diffs back to OOXML in-place. Exposes just 3 tools (read, search, edit) and claims 3x faster, 2x cheaper, and more accurate than python-docx or Office CLI MCPs. Free tier: 500 edits/month.

Anthropic released Claude Sonnet 5.5, an upgrade over Sonnet 5 that runs about 30% faster and costs up to 30% less for most workloads. Positioned as a drop-in improvement for agents doing coding, reasoning, and tool use at lower latency and price.

Claude Code CLI 2.1.283 ships gateway hint headers with an `x-claude-code-prompt-id` for grouping requests per user prompt, new `availableModelsMatch=exact` and `deniedModels` managed settings to lock down model access, a `/doctor prompt-audit` command that scans CLAUDE.md/skills/agents/commands for outdated prompting patterns, MCP tool/WebFetch/WebSearch output capture in OpenTelemetry spans, and many MCP, plugin, sandbox, and vim-mode fixes.

DSPy 3.4 adds Jev integration via the TypeSafe client with three experimental decision types—Noul (boolean), Choice, and Score—that expose probability evidence directly in DSPy signatures. A new ReAnchor optimizer calibrates thresholds/cuts/weights against a program metric. Native lm15 engines replace LiteLLM by default, with `register_provider` for custom OpenAI/Anthropic-compatible HTTP backends, plus async ReActV2 and a local CPython interpreter for trusted code.

Ollama 0.40.0 automatically runs supported model architectures on Apple's MLX runtime on Apple Silicon, delivering better native performance. New MLX-enabled models include gemma4, qwen3.5/3.6/3.8, decision models (Nimble, tev1, clef, clef-flash), and the embeddinggemma-2 embedding model.

Claude Code CLI v2.1.281 expands the Claude apps gateway with AWS Bedrock STS `assume_role` upstreams, per-request Amazon Bedrock guardrail enforcement, telemetry resource attributes, and new desktop policy blocks. It also adds MCP URL-mode elicitation on the 2026-07-28 protocol (servers can open browser flows), MCP checks in `claude plugin validate`, an auto-mode recommendation in `/insights`, plus dozens of resume, retry, sandbox, and permission fixes.

Anthropic's Claude Agent SDK for Python 0.2.158 introduces a `verbatim_prompts` option on `ClaudeAgentOptions`. When enabled, user messages are passed to the CLI as-is with no `@path` file expansion or slash-command dispatch, mitigating prompt-injection attacks where untrusted text could trigger file reads or command execution. Also bumps bundled Claude CLI to 2.1.280. Requires CLI 2.1.248+.

Exa published version 3.4.1 of its official MCP server to the Model Context Protocol registry. It exposes Exa's neural web search, page content retrieval, and multi-step research workflows as MCP tools so any MCP-compatible agent (Claude Desktop, Cursor, etc.) can query the live web and reason over results.

openai-python 3.19.0 adds API support for the new GPT-Rosalind research model and GCP external storage, plus fixes for queued WebSocket omission markers, safer request retries, aiohttp compatibility, and async transform propagation. Agents built on the official Python SDK can now target these new endpoints without custom HTTP glue.

The official OpenAI Python SDK v3.18.0 introduces new model identifiers for GPT-6 Sol and GPT-6 Luna, making these next-generation models addressable through the standard client APIs (chat, responses, etc.).

OpenAI announced GPT-6 in two variants: Sol, a high-capability frontier model, and Luna, a cost-optimized sibling. Both target everyday work with different tradeoffs between raw intelligence and price/latency, giving agent builders a new choice on the OpenAI model shelf.

LiteLLM 1.102.0 adds opt-in per-server OAuth relay discovery for MCP, applies key/team guardrails to MCP tool calls, ships Mistral /v1/audio/speech text-to-speech, adds Vertex Lyria model support, runs post_call guardrail pipelines (with text rewrites) on streaming responses, adds GovCloud Bedrock pricing, and improves adaptive/auto router behavior. Docker images are now cosign-signed.

Claude Code CLI v2.1.280 ships Claude Opus 5.5 (`claude-opus-5-5`) as the new default Opus model with a 1M-token context and $4/$20 per Mtok pricing ($0.20/Mtok cache reads). Also adds `CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH` to override the 2,048-char cap on MCP tool descriptions/server instructions, richer hook OpenTelemetry events, and dozens of fixes to auto mode, subagents, MCP reconnection, and dialog handling.

The Anthropic Python SDK 1.8.0 adds API support for the new claude-opus-5-5 model, inline tool definitions, and a beta MCP tool-list pinning feature that lets clients lock the set of tools discovered from an MCP server. Also includes fixes: add_tools() now takes effect immediately, tool-runner compaction requests omit reply-only params, shared evaluated_permission enum for Managed Agents events, and a Python 3.13 stream-at-exit crash fix.

vLLM 0.30.0 lands 762 commits: new model support for DeepSeek-V4.1-Flash, DeepSeek-V4-Flash-Vision-Exp, GLM-5.3-Flash, K2-Horizon, Cohere Compass and Bailing V3 VL; a persistent IPC weight-cache daemon (`--load-format ipc_cache`) for fast engine restarts; Gumbel-max watermarking with detection; HiSparse host-tier KV offload; Model Runner V2 dual-batch overlap; PCP+DCP large-scale serving; and NVFP4/AutoRound quantization upgrades.

Anthropic announced Claude Opus 5.5, which reportedly matches Claude Fable 5.1 quality on most tasks while running 40% cheaper than Opus 5. Positioned as a more cost-efficient flagship model for agentic and general workloads.

Claude Code 2.1.277 introduces AGENTS.md as a fallback for CLAUDE.md (configurable under Project instructions in /config), a CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1 mode that hands hostnames to a forward proxy instead of resolving locally, and optional static `headers:` on gateway upstreams. Also: many fixes for -p/SDK sessions, resume/cache correctness, plugin install, sandbox globs, subagent MCP re-render breaking prompt cache, and headless cost totals.

Anthropic's Python SDK v1.7.0 adds a compact_before_next_turn() hook to the tool runner, letting agents compact conversation state before the next model turn. The release also introduces a `group` field with display_name on rate limits (deprecating group_type), Bedrock eventstream error handling, and improved parsing of X|Y union types in responses.

OpenAI Python SDK v3.15.0 adds agent session model settings, audio-mini model choices, compaction progress events, managed Responses WebSocket sessions, and prompt-cache prewarming. Also preserves chat stream moderation results and clarifies SIP call ID usage for realtime voice agents.

The official OpenAI Node SDK v7.18.0 adds agent session model settings, audio-mini model choices, managed Responses WebSocket sessions, and prompt-cache prewarming. It also validates WebSocket results, preserves header defaults, and clarifies incoming SIP call ID usage. Meaningful upgrades for building realtime and voice-enabled agents on the Responses API.