Co-op Anything is a playbook plus MCP server that lets an agent detect a single-player game's engine, check a directory of 208 existing co-op projects, and decide between fixing an existing mod or reverse-engineering from zero. Install: `git clone`, then `claude mcp add coop-anything -- python tools/mcp/coop_mcp.py`, or just paste the repo URL to your agent. It solves the problem of wanting to play a single-player game with friends when no co-op mod exists and you have no idea where to even start reverse-engineering it.
Today's drill is the 'implement an ML algorithm from scratch' round: the real skill isn't memorizing formulas, it's vectorized thinking (broadcasting instead of nested loops), numerical stability (subtract the max before exponentiating so softmax doesn't overflow on extreme logits), and edge cases (empty clusters in k-means, tied scores in AUC) — those are where you actually earn points. You're usually restricted to NumPy and expected to write the model, loss, or layer yourself instead of calling sklearn. Today's practice question is adapted from PracHub's machine learning engineer question bank: k-means clustering from scratch.
Haystack v3.4.0, three things worth knowing: (1) new feature — the experimental `ToolResultOffloadCompactor` lets `CompactionHook` write old tool-call results to a `ToolResultStore` and leave only a reference behind; a dedicated read tool can pull the full content back on demand, so nothing is actually lost; (2) breaking — the default embedding model on `OpenAITextEmbedder` and friends changes from `text-embedding-ada-002` to `text-embedding-3-small`; the vectors aren't compatible, so existing corpora need re-indexing or an explicit pin to the old model; (3) breaking — `ChatMessage.from_openai_dict_format` now raises `ValueError` for content parts it can't convert, such as image URL references or `file_id` file references, instead of silently stuffing them into a text string. The same release adds official Python 3.15 support and drops Python 3.10, which reached end of life.
Confido Health closed a $12.5M Series A extension led by Z47, bringing its total Series A to $22.5M. The signal: healthcare admin agents don't grow by selling to clinics one at a time — they grow by selling to the private equity funds that own the clinics, then landing the whole portfolio in one motion.
Muse Spark Open: weights released 2026-10-15, an 816B total / 42B active MoE with a native multimodal Perception Encoder and a 1,048,576-token context window. Third-party API pricing is $1.25 input / $4.25 output per 1M tokens ($0.15 cached input) — identical to the closed Muse Spark 1.2 from two months earlier. In Meta's own comparison table, Muse Spark Open tops every listed model on JobBench (61.8) and MCP-Atlas (89.2), but scores only 39 on the AA Intelligence Index, behind GLM 5.3 (45) and Kimi K3 (44). It's also the first model of its size to run fully on two Apple M5 Ultra Mac Studios via MLX.
Hy-Memory-30B-A3B: open-sourced 2026-10-15 under Apache-2.0, a 30B-total / ~3B-active MoE with a 262,144-token context window, trained specifically to run the "write" path inside agent memory frameworks (Mem0, MemOS, Hy-Memory). Two-stage training — On-Policy Distillation plus Attributed Memory RL — raised the cross-harness macro average from 47.66 (its own base model) to 55.49, which Tencent frames as beating Claude Opus 5's 54.62. Broken down, though, the Mem0/Hy-Memory/MemOS margins are all under 0.6 points — essentially statistical noise — and the real gap comes from the single held-out EverMemOS harness (56.89 vs. 54.04, a 2.85-point difference), which accounts for over 80% of the overall lead. There's no quantized build and no hosted API yet, and the companion MemHub desktop app's install path isn't working end to end.
OpenAI told developers on October 14 that GPT-4o, GPT-4o mini, GPT-4.1, and GPT-4.1 mini — plus every fine-tune built on them — leave the API on January 14, 2027. The named replacement is gpt-6-luna. On list price, GPT-4o's input rate drops from $2.50 to $0.10 per million tokens (down 96%) and output from $10.00 to $0.50 (down 95%). But gpt-6-luna is a reasoning model: at its default medium effort it generates reasoning tokens billed as output, and Chat Completions only supports function calling when effort is set to none. The real savings end up smaller than the sticker price suggests.
Today's practice is the growth PM answer spine: clarify whether 'growth' means acquisition, retention or revenue, map the funnel with AARRR and check whether the product grows through a loop, segment before diagnosing, rank experiments with ICE, and pair the primary metric with a leading indicator and a guardrail. The practice question is Suno's retention drop-off; the case is Calendly's viral loop — scheduling is a problem both sides hit at the same moment, a condition most B2B products can't borrow.
Boundscope is a local C/C++ memory-error testing tool with both a CLI and an MCP server. Install: `cargo build --locked` (no crates.io or npm release yet). It solves the problem of an agent writing C/C++ that compiles and runs fine, with no signal about buffer overruns or use-after-free until something breaks later.
Workday's acquisition of agent-identity startup C1.ai, the AWS Bedrock AgentCorruption exploit chain, and CrowdStrike's disclosure of AI-agent-assisted breaches at South Korean banks all pulled the spotlight from model capability to execution governance; the EU, Taiwan, and Singapore all demanded evidence from agents in the same week — first-ever EU AI Act enforcement, Taiwan's FSC advisory on agentic AI in finance, and Singapore's IMDA testing toolkit; bugs in the AppWorld/WorkArena verifiers, an unreverted database in τ-bench's scoring path, and three leaderboard reshuffles in 72 hours all showed that evaluation infrastructure itself can be unreliable; Meta shipped Muse Wonder/Wonder Full and the open-weight Muse Spark Open in the same week; and Modal's $750M Series D pushed its valuation to $15.75B, cementing agent sandboxes as a standalone business.
Today's practice is the architecture call on splitting agents: every hand-off is another step that can fail, and a 0.95 per-step success rate compounds to about 0.60 over 10 steps, so multi-agent designs only pay off when each worker's scope is narrow; keep tools few and well-scoped, because every tool definition costs context on every request and adds one more wrong choice; MCP connects an agent down to tools and data while A2A delegates long-running tasks between agents, and agents in one codebase should just use the framework's own orchestration; MCP tools, resources and prompts are controlled by different parties, and hints like readOnlyHint are not authorization — the server must enforce it. The practice question is adapted from systemdesign.academy's multi-agent system design problem.
Arena Agent leaderboard published 10-13: Claude Opus 5.5 (Max) stays first at 17.01%; GPT-6.1 Sol (Max) second at 15.05%; Claude Fable 5.1 (Max) climbs from sixth back to third (11.74%→14.32%); GPT-6 Sol (Max) falls from third to seventh (12.60%→10.51%). Three releases in one week, three changes in third place, and overlapping 95% confidence intervals across the top six — the ranks themselves shouldn't drive model selection.
Three things in Mastra 1.77.0: Factory gets a formal sandbox provider contract (Docker, E2B, Mastra Platform) and prebuilt environments; the previous release's run fencing now covers Cloudflare DO/D1, DynamoDB, MongoDB, Redis and a dozen more, so a durable run that lost ownership can't overwrite the recovered run's state; spans support links, and `@mastra/mcp` connects client tool-call traces and server request traces in both directions. Breaking: `@mastra/server` workflow routes can no longer start or resume internal workflows, so durable agents must go through agent routes.
Pydantic AI v2.56.0: `SmolSandbox` runs harness capabilities in a local smolvm microVM; `PGTaskDurability` provides durable execution on pgtask, alongside a Render Workflows integration; `Skills` can layer project and personal libraries; Python 3.15 and Gemini 4 Argon are supported, with the Google providers' default timeout raised to 20 minutes. Compatibility note: `BrowserUse` now runs on the host model by default and browser-use telemetry is opt-in; `MCPToolset(http_client=...)` now applies auth and no longer closes the user-supplied client.
Fragment announced a $12M Series A led by FirstMark with BoxGroup participating (Fortune describes the two as co-leads); the round actually closed in mid-2025. The signal: once agents touch money, 'reconcile before you pay' becomes its own infrastructure layer.
Modal closed a $750M Series D led by Accel at a $15.75B post-money valuation, 3.4x its $4.65B Series C in May. The signal: isolated environments where agents run the code they write have gone from an add-on to a standalone $300M-a-year business.
Zenzap disclosed a $33M Series A co-led by Bessemer Venture Partners and Lightspeed Venture Partners, bringing total funding to about $44M; the round closed in August 2025. The signal: the next entry point for agents may not be a personal assistant but a team work chat that replaces the WhatsApp group.
Today's practice is defining quality for AI products: an AI PRD adds six sections to a normal PRD, and its eval plan is the acceptance criteria; evals start with error analysis on real traces, ranked by frequency times customer cost, with evals written only for the top few errors and the cheapest check (a code assertion) tried first; launch thresholds are written before you see results and split into zero-tolerance and acceptable-rate classes. The practice question is a Perplexity question from the AllthingsPM bank; the case is Intercom Fin moving its billing unit from resolutions to outcomes.
AgentConfig Doctor is a static scanner for agent configuration. Install: from source for now with `npm ci && npm run build` (not yet on npm). It solves the problem that files like `.mcp.json` and `.claude/settings.json` go unreviewed, and you can't tell when they've been changed.
Today's ML System Design rotation covers four core concepts: a feature store needs both an event time and an available time on every feature version to guarantee point-in-time correctness — filtering on event time alone still leaks future information; a serving system should carve its latency budget into explicit sub-budgets (lookup, scoring, response) and version the model, feature schema, and fallback path together; data drift (the input distribution shifted) and concept drift (the input-output relationship itself shifted) need different detection timing — the former shows up in feature distributions immediately, the latter usually needs labels to confirm; and production monitoring splits into three layers — operational (latency/error rate), data (feature distribution/missing rate), and model (prediction distribution/actual accuracy) — where the first two layers serve as early warning when label delay is long. The practice question adapts a real scenario from an Amazon MLE interview guide: a feature is only point-in-time correct when both its event time and its available time are no later than the prediction time, and the exercise is learning to state that criterion precisely in a system design interview.