Your agent says it's done. Check.

MCP went stateless — remote servers just got easy

Source: Claude Blog

On July 28 Anthropic finalised MCP 2026-07-28, the largest change to the protocol since launch, replacing 2025-11-25. The core is now stateless: the protocol moved from a bidirectional, state-holding connection to plain request/response, which means an MCP server can sit on serverless or edge infrastructure without contortions. Authorisation now lines up with how OAuth 2.0 and OIDC are actually deployed, so your server talks to identity systems like Entra or Okta directly. A versioned extensions framework arrived alongside it: MCP Apps for server-rendered UIs inside a sandboxed iframe, and Tasks for long-running and async work. Anthropic puts SDK downloads past 400 million a month — four times higher across the year.

Your takeaway: If you run an MCP server, read what breaks before you upgrade. And if you shelved building one because hosting it was a chore, that excuse is gone — stateless means an ordinary serverless function.

OpenAI introduced Astra with proofs, not claims

Source: GitHub — openai/ten-proofs

On August 1 OpenAI published ten solutions to open problems in mathematics and theoretical computer science, mentioning in the third paragraph that an internal version of Astra — its next major model family — produced them. The problems had gone at least a decade without progress and span group theory, high-dimensional geometry, coding theory, quantum complexity and lattice cryptography. For a builder the result matters less than the shape of the evidence: a 249-page manuscript shipped with Lean 4 certificates for every result, on GitHub under Apache-2.0, so anyone can machine-check them instead of believing them. Total inference cost was about $2,000 at Sol rates. Astra is not a product yet — no price, no context window, no date.

Your takeaway: The lesson isn't about mathematics. Look at what they attached: something a machine can verify. Ask the same of your agent's output — what here can be checked automatically? A passing test, a validated schema, a green build. If there's nothing, you're believing, not verifying.

HUMAIN invests in MOZN — its first bet on a Saudi company

Source: Arab News

On August 3 HUMAIN, the PIF-owned AI company, announced an investment in Saudi firm MOZN plus a partnership to build enterprise AI — its first investment in a Saudi company. The deal pairs HUMAIN's infrastructure with MOZN's products and its Forward Deployed Engineering team, aimed at financial institutions and government bodies. MOZN already serves more than 150 customers across those two sectors. First solutions arrive at LEAP in Riyadh, with wider commercial availability in the second half of 2026. The stated goal is moving institutions off pilots and into secure production.

Your takeaway: Read this as a market signal, not a funding item. Buyers in regulated sectors want something that runs inside their borders — data sovereignty, and engineers who sit with them. If you sell to Arab banks or government, those two things close the deal, not your feature list.

A breach in the Middle East now costs $8 million

Source: IBM / Zawya

IBM's annual Cost of a Data Breach report, out this week, puts the average breach in the Middle East at $8 million. Twenty-six percent of malicious breaches involved AI, and another 11% of respondents couldn't say either way. The three biggest cost drivers in the region: mismanaged secrets and keys, excessive privileges and poor role management, and an inability to prioritise threats. The uncomfortable part — organisations using AI and security automation extensively averaged more than $3 million lower, and 23% still haven't adopted either.

Your takeaway: Start at the top of that driver list: secrets and keys. Open your repo now and go looking for an API key in the code or in git history. The region's most expensive breaches start there, not with a clever attack.

open-code-review — Alibaba's reviewer is climbing GitHub fast

Source: شُهُب

Alibaba's open-source code review tool was among the fastest-growing repositories this week: 18,698 stars after gaining 3,881 in seven days — roughly 238 a day since it appeared in May. It blends deterministic checks with an LLM agent to produce line-level comments, and ships a multi-language ruleset covering NPE, thread-safety, XSS and SQL injection. Written in Go, Apache-2.0, with GitHub Actions and GitLab CI integration.

Your takeaway: Run it over your oldest open PR before you read that PR yourself, then compare. What did it catch that you'd have missed? If the answer is nothing, you saved an afternoon. If it's one real thing, it already paid for itself.

i-have-adhd — stops your agent burying the answer

Source: GitHub

A small Python skill aimed at a complaint anyone working with a coding agent knows: the answer is in there, buried under paragraphs of explanation. The repo has reached 16,617 stars and 933 forks since May — about 201 a day — under MIT. The idea itself is plain: shape the output so the answer comes first.

Your takeaway: Cheapest experiment on this list. Install it, then ask your agent the same question that irritated you yesterday. If the answer lands on line one, you've won. The difference isn't model intelligence — it's output formatting, and that's yours to control.

openwork — an open alternative to Claude Cowork

Source: GitHub

A TypeScript project positioning itself as the open-source alternative to Claude Cowork, built on opencode. It's at 20,844 stars and 2,046 forks, up 3,429 this week, with commits as recent as August 4. One flag before you build on it: GitHub doesn't recognise its licence as a standard one (NOASSERTION) — read the licence file yourself before any commercial use.

Your takeaway: Try it on a side project, not your product. A non-standard licence isn't necessarily a problem, but it is a question — and the time to ask is before the thing sits at the centre of your business, not after.

LongHorizon-Harness — get task state out of the context

Source: arXiv

Eight researchers posted a sharp diagnosis on arXiv (2608.01964) of why agents fail at long tasks: existing harnesses keep execution, task state and completion assessment all inside a context that keeps growing, so state gets lost and the agent's own faulty self-assessment leaks into later decisions. Their fix is structural — hold task state outside execution, and update it only with facts independently verified from the environment. The loop is Manage-Execute-Audit: a manager holds state and picks the next subtask, a fresh-context executor does it, and a read-only auditor confirms the environment state before the next round.

Your takeaway: They didn't change the model — they changed where state lives, and success went from 51.8% to 80.7%. If your agent stumbles on long tasks, the model probably isn't the problem. Move state somewhere you write, and update it only from things you can actually confirm: a file exists, a test passed, a service responded.

SWE-Touch — what happens when you touch the code too

Source: arXiv

Seven researchers (arXiv 2608.02499) asked a question familiar to anyone who codes alongside an agent: what happens when a human edits the code mid-task? Existing benchmarks test the agent alone, or limit user participation to messages. So they built a framework that injects "counter-edits" — plausible changes to task-relevant code that conflict with finishing the task — and delivers them with a contextual message when the agent reaches that spot. Across nine coding models, resolve rate fell 7.7 percentage points on SWE-bench Verified, and the drop held on longer tasks in SWE-Bench Pro and DeepSWE.

Your takeaway: If you edit files while your agent is working in them, you're paying 7.7 points of its accuracy without noticing. The practical rule is simple: split ownership. Either the file is its or yours — and tell it plainly what you changed before it carries on.

Move your MCP server to stateless

Source: Claude Blog

The new spec removed the heaviest part of running one: state. Your server can now live as an ordinary serverless function with no open connection, and its authorisation plugs into standard OAuth/OIDC. If you have an MCP server carrying state between requests, this is a small rewrite that buys cheaper deploys and scaling you don't have to think about. And if you don't have one, building it today is simpler than it was last week.

Your takeaway: Start with your smallest tool — one that reads and never writes — and ship it stateless. When that works, move the rest. Don't rewrite everything at once.

Build it and launch it on Shuhub →

Give every agent output a certificate someone can check

Source: GitHub

Three items in this issue say the same thing: Astra shipped Lean certificates with its proofs, LongHorizon-Harness updates state only from environment-verified facts, and SWE-Touch showed agents don't notice an edit that contradicts them. The opening: instead of reading your agent's summary, require it to leave behind something a machine can check — a passing test, a validated JSON schema, a successful build. Make that an acceptance condition in your pipeline, not a personal habit.

Your takeaway: Write the rule down in one place: no agent's work ships without something else able to verify it. One line in CI enforces that better than ten manual reviews.

Build it and launch it on Shuhub →

Audit your own secrets before someone else does

Source: IBM / Zawya

IBM put mismanaged secrets and keys at the top of what drives up breach cost in the region. This isn't a big-company problem: an API key in an env file committed by accident, or sitting in git history, is enough. The opening for an Arab builder is double — clean your own house first, then notice that every secret-scanning tool assumes a team that reads English and follows policies written in it.

Your takeaway: Give it an hour today: run a secret scanner over your repository's entire history, not just the latest commit. A leaked key isn't deleted when you delete the file — it stays in the log until you revoke it.

Build it and launch it on Shuhub →