Radar · 2026-08-14

Generated by a daily research loop, unedited. 14 items. Evidence is quoted from the scan; links go to the source.

Trending repos

cloudflare/computer — a filesystem + exec runtime for agents inside a Durable Object

[Cloudflare changelog, Aug 3 2026](https://developers.cloudflare.com/changelog/post/2026-08-03-cloudflare-computer/) — SQLite-backed virtual filesystem in a Durable Object, with three interchangeable backends behind one `workspace.runtime.exec()`: Workers isolate JS, `just-bash` isolate shell, and full Linux container. +3,599 stars in 7d (#2 TS trending). Ships a Computer MCP example. Marked PREVIEW, unstable API.

TencentCloud/TencentDB-Agent-Memory — team-scoped agent memory with L0→L3 layering

+5,388 stars this week (21.5k total), v2.0.0 "Team Memory Beta." Architecture: L0 raw conversations → L1 atoms → L2 scenarios → L3 personas; retrieval is BM25 + vector with RRF; four asset types (Chat Memory, Skills, Wiki, CodeGraph) with per-agent ACLs. Notably *no* MCP server — exposes `/v3/tools/list` and `/v3/tools/call` instead.

Meta Muse Glimmer 30B — open-weight agentic model built for local tool calling

[HN #1 this week, 1,202 pts / 638 comments](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model). Apache 2.0, 128K context, four reasoning-effort levels, schema-based tool calling and failure recovery as design goals. [Live in the Ollama library](https://ollama.com/blog/muse-glimmer) (needs 0.32.7+); GGUF quants down to 2-bit with imatrix calibration.

semantica-agi/semantica — polyglot graph-native context infrastructure

+4,073 stars this week (#1 Python trending). Backend-agnostic across Neo4j/FalkorDB/Apache AGE *and* pgvector/Qdrant, plus RDF stores. Ships a full MCP server (entity extraction, decision recording, graph queries, reasoning). At v0.6.5, a security release patching API auth, injection, and SPARQL validation — young code.

vitali87/code-graph-rag — knowledge-graph RAG over a monorepo

+1,628 stars this week (#5 Python trending). Builds a multi-language code knowledge graph for cross-repo querying rather than flat chunk embedding.

Releases worth your attention — week of Aug 7–14, 2026

Vinext is now the default Next.js scaffold in create-cloudflare

`[email protected]` (Aug 13) — "Vinext is now the default Next.js framework option with OpenNext as an alternative." Vinext reimplements the Next.js API surface on Vite; ~94% of the Next.js 16 surface, 4.4x faster builds, 57% smaller bundles. Cloudflare flipped the default, demoting OpenNext to opt-in.

Mid-conversation tool changes beta — add/remove tools without nuking the prompt cache

`anthropic-sdk-python` v0.121.0 (Aug 7) adds the `mid-conversation-tool-changes-2026-07-01` beta, plus session budgets, an advisor tool, pinned inference location, and skills auto-loading from GitHub. Declare the full tool set up front, then use `tool_addition`/`tool_removal` blocks on `role: "system"` messages — both support `cache_control`. Previously, editing `tools` invalidated the cache for the whole conversation.

`output_behavior` on dream creation — in-place memory store updates

Anthropic TS SDK v0.117.0 and Python v0.122.0, both Aug 13: "add `output_behavior` to dream creation (create a new memory store or update the input store in place)." Dreaming was previously strictly copy-on-write — a dream consumed a memory store plus up to 100 sessions and emitted a *separate* store. In-place update is genuinely new.

Ollama v0.32.9–v0.32.11 — Nemotron 3.5 Lightning, web search in Responses API

Three releases Aug 11–14. v0.32.9 adds NVIDIA Nemotron 3.5 Lightning, "an open 30B MoE with 3B active parameters built for that execution layer of always-on agents." v0.32.11 (Aug 14) adds web search to the OpenAI-compatible Responses API. v0.32.10 silently changed default `repeat_penalty` from 1.1 → 1.0 for models that don't specify it.

Tech-direction conversations

MCP v2: stateless spec, DCR deprecated, no more Durable Objects

[blog.cloudflare.com/mcp-v2](https://blog.cloudflare.com/mcp-v2/) (Aug 6, headline of Agents Week; [recap Aug 10](https://blog.cloudflare.com/agents-week-review-august-2026/)). Handshake/session management removed — each request carries protocol version + client identity. New `Mcp-Method`/`Mcp-Name` HTTP headers let WAFs and rate limiters route without parsing JSON. Dynamic Client Registration deprecated in favor of Client ID Metadata Documents; Roots, Sampling, Logging and HTTP+SSE deprecated with a 12-month minimum support window. Tool lists now carry `ttlMs`/`cacheScope`.

Cloudflare AI Search — managed RAG with a built-in `/mcp` endpoint

[blog.cloudflare.com/ai-search-easier](https://blog.cloudflare.com/ai-search-easier/) (Aug 6). Collapses Workers AI + Vectorize + R2 + AI Gateway into one product; exposes `/search` and `/mcp`. Embedding and reranking are free on default models; ingestion $0.75/1M tokens (5M free), storage $2/GB-mo (10GB free), queries $0.75/1k.

Meta's Muse Glimmer: 30B Apache-2.0 model built for always-on local agents

[research.meta.ai](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) (Aug 10) — HN #1 of the week at **1202 points / 638 comments**. 4-bit quant lands under 20GB, targeting M4/M5 MacBooks and 5090s. Ships a drafter model for 1.5–3.1× speculative decoding, tuned for tool invocation and failure recovery, Apache 2.0.

DeepSeek Harness: MIT agent framework with append-only, forkable session logs

[deepseek.com/harness](https://deepseek.com/harness/en/) (Aug 13, 667 pts / 276 comments). Everything is a swappable plugin — models, tools, sessions, sandboxes, storage, loops. Every token the model sees goes into an append-only session log you can inspect, resume, **fork, search and replay**. Includes a "code mode" where the model orchestrates via TypeScript rather than tool JSON.

Agent sandboxing consolidates on microVMs — and the network is the gap

[Docker Sandboxes](https://www.docker.com/products/docker-sandboxes/) (Aug 10, 692 pts / 397 comments) ships free disposable microVMs designed for `--dangerously-skip-permissions`, supporting Claude Code, Codex, Gemini CLI. Two days later Vercel published ["A sandbox without a network boundary is only half a sandbox"](https://vercel.com/blog/a-sandbox-without-a-network-boundary-is-only-half-a-sandbox) (Aug 11), arguing filesystem isolation alone leaks via egress.