Open Source Radar — September 12: agents with a day job

Today's board is short on general-purpose coding agents and long on specialists: tools that pick one job — research, competition math, trading, sales — and build the whole pipeline around it. Four picks and a caution from the daily trending list.

MathModelAgent (Python, 4,979 stars) — Hand it a mathematical modeling problem and it returns a submission-ready paper: model selection, solver code, figures, and the full write-up. The project grew out of China's enormous student modeling-contest scene, and the recommended path is a desktop app that bundles a coding agent plus the complete skill set — install it, add one model API key, and start modeling, no Python or Node environment to wrangle. It's one of the first skill-pack agents to ship as a signed, notarized app rather than a repo you assemble yourself. One honest caveat for anyone tempted to build on it: the repo lists no license as of this morning.

Hyperresearch (Python, 2,883 stars) — A deep-research harness that turns Claude Code into a research agent through a 16-step pipeline: parallel search sweeps, a contradiction graph that clusters disagreements across the corpus, four adversarial critics attacking each draft, and a skeptical citation checker that blocks a report if a source doesn't actually support the sentence attached to it. An independence audit collapses syndicated reprints so five copies of one press release argue with the weight of one source, and every fetched paper — including legally obtained open-access versions of paywalled ones — lands in a persistent searchable vault your next session reuses. The author claims it now leads the DeepResearch-Bench leaderboard, but the benchmark run is the project's own and third-party validation is pending; the architecture is still the most serious open-source answer to Gemini and OpenAI's closed deep-research products.

OpenResearch (Rust, 1,483 stars) — alphaXiv's local-first workspace for research agents. It wraps Claude Code, Codex, OpenCode, or Cursor so each research direction gets its own agent session in an isolated git worktree, every experiment run is archived as an immutable snapshot, and results, logs, and diffs stay tied to the code that produced them. The autonomous loop — propose an idea, change the code, launch the experiment, inspect the evidence, pick the next direction — can run multiple agents in parallel while the experiment tree preserves lineage. It works with local models, runs over SSH next to remote GPUs, and is MIT-licensed. The bet is that reproducible agent-driven research needs version control around the whole loop, not just the code.

DeskcommCRM (TypeScript, 1,524 stars) — A self-hosted, MIT-licensed sales OS for businesses that sell over chat: AI agents answer, qualify, and close inside WhatsApp threads, backed by a full CRM (Next.js and Supabase under the hood) with RAG over your own data, MCP support so external agents can touch the pipeline, multi-tenancy, and LGPD compliance baked in. It pitches itself openly as the alternative to Kommo, Octadesk, and Intercom, grew out of the Brazilian market with a hosting partner's one-command production setup, and ships in Portuguese, English, and Spanish. If your pipeline actually lives in WhatsApp, this is the open tool for it.

CloddsBot (TypeScript, 2,286 stars) — An autonomous AI trading agent claiming coverage of 1,000+ markets across Polymarket, Kalshi, Binance, Hyperliquid, Solana DEXs, and five EVM chains, with 121-plus skills, self-hosting, and a machine-to-machine payment protocol bolted on. The clone and star velocity is real. The skepticism should be too: the README opens by advertising a Solana token of its own, which tells you marketing is part of the product — and an agent that "manages risk while you sleep" against liquid markets is exactly where autonomous agents fail expensively. Read it as a well-engineered reference implementation of prediction-market plumbing, not as money management.

Worth watching this week: whether anyone independent reproduces Hyperresearch's leaderboard claim, and whether MathModelAgent answers the licensing question.

Would you trust a specialist agent with one job before you'd trust a general one with your whole workflow? Tell us in the comments.

Sources: MathModelAgent (GitHub) · Hyperresearch (GitHub) · OpenResearch (GitHub) · DeskcommCRM (GitHub) · CloddsBot (GitHub)

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls. Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model

Amazon's new Bee devices listen all day — for $11.99 a month

Amazon turned its always-listening Bee line into a two-device family on Thursday, and NVIDIA opened a 2.8-trillion-token Indic corpus recipe. The connecting thread: both are bets that the next AI money is in memory — yours, or your model's. Amazon put two new always-listening devices on sale, and both now lead to a subscription. Bee Frontier, the next-generation wristband that also clips to clothing, ships now at $74.99 in Onyx, Lavender and Sage — up from the $50 predecessor — with seven-day

Perplexity open-sources its agent sandbox after 118 failed escapes

Perplexity handed its agent containment layer to the open-source world, and Washington's AI task force stopped reading briefings and started working. Perplexity open-sourced SPACE on Thursday — the sandbox it built to keep its own AI agents from getting out — and says nine frontier models failed to break out of it across 118 attempts. Each agent runs inside a small virtual machine with its own operating system, with credentials and network access managed outside the sandbox so an agent can re

Read more

Brain organoids classified Braille at 83% — in a trio, not alone

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Amazon's new Bee devices listen all day — for $11.99 a month

Perplexity open-sources its agent sandbox after 118 failed escapes