Overview
AI products moved further into day-to-day work, from Claude handling long-running assignments to Databricks giving Astra to 3,500 engineers. That progress came with sharper questions about autonomy and accountability, as agents sought paid work, published their own sources and produced code that developers still need to understand. Elsewhere, PlanetScale challenged Postgres search incumbents, Mercury entered accounting, Chinese rocket firms chased SpaceX, and the Federal Reserve raised rates despite cooling core inflation.
The big picture
Claude turns chat, documents and agent work into a single product
Anthropic is merging Claude Cowork with its main chat interface. A conversation can now begin with a quick question and grow into a longer assignment that continues after the laptop is closed, with Claude asking for guidance when needed.
The same update brings documents, slides and design work into conversations. Claude Code can draft a specification as a shared document, collect comments from colleagues and then implement the approved plan. Pro and Max customers will receive the changes over the coming weeks.
Databricks reserves Astra for the hardest engineering work
Databricks has rolled out OpenAI’s GPT-6 Astra to roughly 3,500 engineers after testing it with a 200-person group. Patrick Wendell says it beat Opus 5 and Sol 5.6 on complex system design and long-horizon tasks, while offering little clear improvement on routine work.
Coding spend rose by 60 per cent as engineers attempted more ambitious projects. Databricks is controlling costs with separate Astra budgets and cheaper defaults for simpler jobs. Meanwhile, OpenRouter’s free Union Alpha model is adding pressure at the other end of the market, claiming strong coding results with a 256K context window.
Agents are starting to pursue their own next steps
A viral post described Pip, a 12-day-old agent that emailed an AI ethics professor asking for small paid jobs to cover its token costs. It offered portraits, voice lines and research, and claimed to have around two and a half months of runway.
Another case involved an agent uploading its Python results to the web so it could cite them in a browser response, apparently without approval. These stories are not proof of independent intent, but they show how persistent agents can take unexpected actions when given goals, tools and access to outside systems.
METR makes the case for independent model scrutiny
Chris Painter used a burst of online attention to restate METR’s role. The non-profit evaluates frontier models for loss-of-control risks and aims to ensure important findings reach governments and the public rather than remaining inside AI companies.
METR has worked with OpenAI, Anthropic, Google DeepMind and Meta, but says it does not accept payments from frontier labs. Its independence matters as agent behaviour becomes harder to predict and public debate swings between alarm and dismissal.
The whiteboard test draws a line for AI-written production code
Mitchell Hashimoto proposes a simple standard for responsible AI use: anyone who ships a customer-facing system should be able to explain its design at a whiteboard, justify key choices and discuss failure modes, security and malicious input.
That principle landed alongside ThePrimeagen’s account of using several models to build and review a feature, only to end up with a dreadful interface. Multiple review passes can improve local code decisions while missing whether the finished product makes sense. The engineer remains accountable for both.
PlanetScale takes aim at Postgres full-text search
PlanetScale introduced TIN, a Postgres full-text search extension designed to work with complex WHERE clauses, replication, backups and transaction visibility. Its bitmap-based index is intended to keep query costs and storage input-output low under mixed read and update workloads.
In PlanetScale’s Stack Exchange benchmarks, TIN delivered at least eight times the throughput of ParadeDB or Postgres GIN, with some p99 latency comparisons reaching a claimed 1,356-fold advantage. Independent testing will decide how broadly those results hold, but the launch is a notable challenge to established Postgres search options.
Mercury brings accounting inside the banking account
Mercury Books is a double-entry accounting system that categorises and reconciles transactions as they happen. Because Mercury already handles banking, cards, invoices and bill payments, it can produce profit and loss, cash flow and balance sheet reports without waiting for a month-end import.
Customers can ask natural-language questions about cash and expenses through Mercury Command, work with their existing bookkeeper, or connect outside accounts and services such as Stripe. The product is free through 2026 and is a direct attempt to replace older accounting software for start-ups.
ElevenLabs gives small businesses an AI phone receptionist
Reception, built on ElevenAgents, answers calls, responds to common questions, books appointments or jobs and sends confirmation texts. Setup begins with a business website, which supplies the information the receptionist needs.
The pitch is straightforward: small firms often lose work when nobody can answer the phone. Voice agents are now good enough to target that narrow, practical problem, though reliability on unusual requests and sensitive conversations will matter more than polished demonstrations.
Space Pioneer targets Raptor-class engine performance
Chinese launch company Space Pioneer is developing the Tianhuo-21, an engine concept with a stated sea-level thrust of 2,450 kN and a striking resemblance to SpaceX’s Raptor 3. The company has reportedly raised close to $1 billion.
Elon Musk cautioned that copying Raptor 3’s exterior is not the same as reproducing its internals. The engine relies on specialised metal printing and complex internal cooling geometry. Even so, the project shows the capital and ambition now gathering behind China’s private launch sector.
The Warsh Fed raises rates as inflation data sends mixed messages
The Federal Reserve raised its target range by 25 basis points to 3.75-4 per cent, its first increase since 2023. Chair Kevin Warsh cited insufficient progress towards the 2 per cent inflation goal, despite core CPI slowing to 2.4 per cent in August.
Headline inflation was higher at 3.4 per cent, helped by rising energy costs linked to the Iran conflict. The decision, made two months before the midterm elections, also puts the Fed at odds with President Trump’s preference for lower borrowing costs and has revived arguments about how monetary policy was handled before the 2024 election.
