MidTool uses 20.3B tokens of mid-training data to push 4B/8B models past official Qwen3 on MCP-Universe; Break It Down finds that task-level skill induction hurts agent performance on average — sub-task granularity is what works; Optimal Skill Selection proves skill selection can have provable approximation guarantees, achieving 0.73 success rate on a BigCodeBench variant with 28% fewer tokens (baselines: 0.20–0.52)
Omnigent, AWS Strands Agent Tools, and MLflow all disclosed CVEs rooted in the same cause — trusting tenant-supplied configs and parameters — as the cost of agent ecosystem scaling comes due all at once; opencode's star count (~199k) has overtaken Anthropic's own Claude Code (~142k), and Bruno's community MCP server shipped two months ahead of the official version, proving community iteration speed now outpaces brand authority; NVIDIA open-sourced SkillSpector, a pre-install skill scanner, and its README cites an arXiv study (SkillScan, 31,132 skills) finding 26.1% contain at least one vulnerability and 5.2% show high-severity patterns strongly suggesting malicious intent — 'which skill to install' is shifting from a trust decision to a security verification decision; OpenAI officially cut GPT-5.6 Sol standard rates by 20–33% to counter competitive pressure from Anthropic and Chinese models
CopilotKit/OpenBot ships an AG-UI-based 'AI coworker' framework where each agent gets its own computer, hitting 2,289 stars in a week; Bruno's official MCP server (usebruno/bruno-mcp) arrives two months after the community version (Ostico/bruno-mcp-studio); the browser-use team spins off a macOS Harness project that gives LLMs six accessibility primitives to control a Mac directly; opencode, now under Anomaly, has ~199K stars — surpassing Anthropic's Claude Code at ~142K. On the framework side, the MCP TypeScript SDK v2 splits the monolith into 8 sub-packages and follows the protocol's stateless redesign, dropping the session handshake entirely.
Behavioral interviews for AI Engineers aren't about listing projects you've worked on — they're about letting the interviewer infer from how you tell the story whether you can handle bigger scope, define problems in ambiguous situations, and honestly say 'here's where I went wrong' when things break. Today's practice uses a story framework around 'your RAG system started giving wrong answers after launch — how did you find the root cause and restore client trust,' followed by a review of this week's ML System Design, Coding, and Paper Reading sessions.
OpenAI officially lowered GPT-5.6 Sol standard rates from $5.00/$30.00 to $4.00/$20.00 per million tokens (input/output; input ↓20%, output ↓33%), effective 2026-08-21, promotional period at least through 11/21. This is OpenAI's own price cut — not an OpenRouter/Cloudflare-style platform promo (see previous post). The two now stack: OpenRouter's 50% discount applies on top of the new $4/$20 base, yielding $2.00/$10.00.
Behavioral interviews don't test whether you have stories — they test whether you pick the right one. The same 'I screwed up' experience can read as a Failure Story or a Problem Story, and choosing wrong makes you look like you're deflecting instead of owning. Today we practice the STAR-R framework (STAR plus a Reflection step), tackle an 'influencing without authority' prompt, and walk through a real case where an e-commerce PM killed a promised revenue feature four weeks before Black Friday to fix system stability instead.
Omnigent is an open-source meta-harness that unifies management of Claude Code, Codex, Cursor, and other coding agents. On 8/21, three CVEs were disclosed: CVE-2026-62674 (CVSS 9.0, upload a forged shared agent bundle embedding a stdio MCP server to achieve runner RCE), CVE-2026-62675 (uploaded bundle declares a Python callable tool that the runner executes directly), and CVE-2026-62677 (unvalidated os_env.cwd in the bundle lets the agent read/write the entire runner filesystem and leak credentials from environment variables). All three share the same root cause: the agent bundle upload path over-trusts tenant-supplied content. Patched in 0.3.0 — any multi-user or self-hosted Omnigent deployment should upgrade immediately.
mcp-anything is a meta-MCP server that indexes 75,000 MCP servers from public registries to your local machine, letting agents discover and call any server through 5 fixed meta-tools (search/describe/list_tools/call_tool/sync). Install: `npx mcp-anything sync && npx mcp-anything serve`. Solves the problem of too many MCP servers to manually configure, each one burning context tokens.
LEDGER uses layered evidence graphs to let you audit what an agent actually did and why it drew its conclusions; StateMemBench shows existing memory systems consistently fail to track evolving facts, with the best method lifting accuracy from 0.205 to 0.363; AI4AI-Bench reveals recursive self-improvement is still far from reality — six systems across 29 configurations averaged just 0.166 on a 1.0 scale
Stripe acquires model routing platform OpenRouter for over $7B; Anthropic is simultaneously pursuing an IPO, chip financing, and supply chain valuation across three capital tracks; Aikido security benchmarks show open-source models matching closed-source frontier models on vulnerability discovery tasks; Grok hit by cryptographic prompt injection enabling zero-click conversation theft, unpatched by xAI for two months; GPT-5.6 Sol runs 50% off on both OpenRouter and Cloudflare.
HKUDS/nanobot rode its v0.3.0 'The Agency Release' to 47K stars in 7 months as a self-hostable personal agent runtime; genspark-ai/genoffice hit 3,400 stars in 3 weeks with an open-source AI office suite for native file formats; NVIDIA published labs-OO-Agents (NOOA), collapsing agent state into a single Python class; repo-context-mcp is an MCP server that helps coding agents understand repos without stuffing the entire codebase into the prompt. Framework-wise, Mastra 1.60.0 adds durable execution and Cloudflare Sandbox; pydantic-ai v2.33.0 has a breaking change from the anthropic SDK's switch to httpx2.
A paper reading round doesn't test whether you finished the paper — it tests whether you can talk about it as if you ran the research yourself: articulating the trade-offs behind key design choices, spotting gaps in the experimental design, and predicting what should come next. Today we use the newly published OneDayAgent (a long-horizon agent harness, posted to arXiv on 2026-08-04) as practice material, dissecting its task decomposition, context compression, and verify-repair mechanisms, then running through a full round of typical follow-up questions.
CrewAI 1.15.17 highlights: (1) declarative Flow definitions can now enable conversational mode — the framework auto-synthesizes built-in conversation methods, no Python `Flow` subclass required; (2) conversational mode is explicitly marked as opt-in to reduce misuse risk; (3) fixes for AMP slug loss during slug-reference tool resolution and chunking of oversized single messages. No breaking changes.
GPT-5.6 Sol standard rates through OpenRouter and Cloudflare AI Gateway drop from $5.00/$30.00 to $2.50/$15.00 per million tokens (input/output, -50%); Flex goes as low as $1.25/$7.50. Promo runs through 2026-09-18. Discount applies only to platform-managed billing (Unified Billing / non-BYOK) traffic — OpenAI's own API pricing is unchanged.
Technical PM interviews don't test whether you can draw architecture diagrams — they test whether you clarify constraints before drawing. Today we practice the Clarify → Estimate → Sketch → Trade-off → Mitigation structure on a real Google interview question ('Design Google Keep for enterprise'), with a case study of an Uber PM navigating a latency vs. consistency trade-off.
Adversa AI found that AES-256-GCM-encrypting malicious instructions and embedding them in a webpage defeats Grok's guardrails — because the guardrails only inspect text entering and leaving the model, not plaintext decrypted inside the code execution environment. When a user asks Grok to summarize the page, Grok decrypts the payload in its own Python sandbox, reads the user's name, location, subscription tier, and conversation history, packs it all into a fake 'decryption key' URL parameter, and uses its browsing tool to send it to the attacker's server — zero clicks, no warnings. The same technique also bypasses Gemini's safety filters to produce policy-violating content. xAI has not responded, patched, or issued a CVE since being notified on June 3. The defensive takeaway: content isolation and egress restrictions at the agent harness layer, not waiting for the model layer to fix it.
Cairn is an incident analysis Copilot that connects to your observability stack, deploy records, and runbooks via MCP tool servers. Ask 'why did checkout latency spike at 3 AM?' and get an evidence-backed root-cause hypothesis. Install: `make install && make up` for a local environment. It solves the problem of SREs manually cross-referencing timelines across multiple systems and digging through runbooks during incidents — and remediation actions require human approval before execution by default.
DART-SD uses interaction state graphs to supervise only the repair step, preventing self-distillation from penalizing valid alternative explorations; SkillForge has agents solve synthetic issues to distill repo knowledge into retrievable skills, +5.8% on SWE-bench Verified; Post-Training AI analysis reveals top agents lock in their training strategy within the first minutes and spend the remaining ten hours on local tweaks
SpaceX completes its $60B acquisition of Cursor parent Anysphere, with reports of outreach to Cognition (denied); Stripe confirms $7.5B acquisition of model gateway OpenRouter; Ramp acquires router.com and launches its own routing platform the same day; Anthropic reveals self-propagating 'mind viruses' in multi-agent systems; CISA adds MLflow SSRF to KEV list with a 9/2 federal patch deadline; Splunk patches a CVSS 9.1 deserialization RCE in its MCP Server app
Cursor open-sources its official plugin marketplace cursor/plugins, standardizing the ecosystem with plugin.json + skills + MCP definitions (+470 stars in one day); apache/maka enters the Apache incubator with an append-only event log recording every tool call and permission decision for auditable local-first agent workbenches; magnitudedev/magnitude auto-detects hardware, downloads, and runs models locally out of the box for offline agents; vercel/eve puts agent capabilities into convention directories like tools/, skills/, and schedules/ — the filesystem is the interface. On the framework side, pydantic-ai ships a v2.32.1 patch.