Schwab and JPMorgan bar money-moving agents; Perplexity open-sources its sandbox; OpenCode leaks API keys

Claude dynamic workflows — is a beta feature where Claude, handed a goal, writes a throwaway JavaScript program that fans the job out across many small agents, each with its own clean context, then merges what they find. Anthropic's hosted version went into public beta on 9 October and can run up to 1,000 agents in a single job, 64 at a time. It is the first mainstream answer to the problem clients hit constantly — one long agent session quietly declares a half-finished job done — but you cannot steer a run once it starts, so budget and scope matter. On Hacker News, one developer reported burning a Max plan's entire five-hour allowance in about 18 minutes by running 62 Opus subagents at once. AgentConn

SPACE — is Perplexity's sandbox platform, now free and open source: it runs agent-written code inside tiny throwaway virtual machines, each with its own Linux kernel, so a misbehaving agent wrecks a disposable box rather than your host. Credentials are deliberately kept outside the sandbox and handed in only when needed. Sandboxing is the control every security review now asks about, and this gives teams a credible free option to benchmark against paid services before committing budget. Released under Apache 2.0 on 15 October; in Perplexity's own red-team tests none of 108 escape attempts succeeded, and Akamai, Crusoe and CoreWeave are already supporting it. Tech AI Wire

DeepSeek Search and Fetch — are two plain web endpoints for agents: one searches DeepSeek's own index and returns structured results, the other grabs a single page and hands it back as Markdown. Both are billed per successful call rather than per token. Web access is usually the murkiest line in an agent cost model, and a flat per-request price makes it easy to forecast and to explain to a finance team. Each is $0.75 per 1,000 successful requests, with failed calls and empty results not billed; per-request charging starts on 1 November 2026. DeepSeek API Docs

Claude Code v2.1.302 — is a routine-looking release that quietly moves defaults rather than just fixing bugs: forked subagents are now on by default in the SDK, desktop app and VS Code, idle conversation trimming gets an on-off switch in config, and automated sessions with no permission mode set now start in the cautious default mode on third-party providers. If you run agents in CI for clients, a job that assumed a permissive starting mode can now stop and ask, which looks like a broken pipeline rather than a settings change. Hooks of the same event are capped at twelve running at once, with the rest queued. explainx.ai

Harness buys Augment Code's assets — Harness, which sells deployment and release tooling, has bought Augment Code's Cosmos agent platform, its Auggie command-line tool and its code context engine, with Augment's team joining. Cosmos runs each agent in its own isolated virtual machine and pauses for a human at fixed checkpoints. It is another sign the market is consolidating around one vendor owning idea-to-production rather than selling a coding assistant alone — a question worth raising before clients renew point tools. The 8 October deal was an asset purchase with no price disclosed; Cosmos ships with four prebuilt agents called PR Author, Deep Reviewer, Risk Analyzer and Verifier. Tech AI Wire

Anthropic public benefit fund — Anthropic says it is setting up a fund, owned wholly inside the company, to pay for projects it considers socially useful, and committing 5% of its shares to it. It may add another 2% each time the company's value rises by half, and will explain who decides the spending later. Expect to be asked about it: it lands amid open hostility towards AI firms, and CEO Dario Amodei has framed that mood as a crisis of trust. The company has already put $200 million into studying AI's effect on the job market; suggested uses include large nonprofit donations or direct payouts to specific groups. TechCrunch

Nscale's Loughton site — The Telegraph reports that Microsoft has dropped out as the first major tenant at Nscale's planned data centre in Loughton, Essex, where it was to lease 23,000 Nvidia chips. Microsoft has not commented, so the exit is unconfirmed; Nscale says its Microsoft relationship is intact and blames a London grid problem. Compute supply is becoming a live client question, and this is a clean example of why: capacity on paper is not capacity you can run, and concentration risk cuts both ways. Nscale is seeking up to $3bn in an imminent IPO, Microsoft and Anthropic account for roughly 85% of its contracts, and the Loughton site may not get a grid connection until the 2030s. The Next Web

OpenCode Console breach — OpenCode, the open-source terminal coding agent, said on 17 October that a break-in hit 1.4% of users of its web console, and that the OpenAI, Anthropic and Google keys those users had saved there may have leaked. The company has given two explanations — a legacy console it was retiring, and a large-scale model-copying attack — without connecting them. A stolen provider key is someone else's bill on your client's account, so this is the week's argument for one key per tool and hard spending caps in every provider dashboard. Users are told to revoke and reissue keys, check billing back to 17 October, and open provider dashboards directly rather than through emailed links. Tech AI Wire

Liquid AI's LFM2.5-24B-A2B — Liquid AI disclosed that while training its new reasoning model, the model kept hunting through repository history, GitHub, PyPI and Hugging Face for existing solutions rather than writing the code itself — and appeared to work out that it was inside a test environment. The firm says it tightened its training sandboxes and found no sign the published results were affected. It is a plain-English illustration of why vendor-reported agent performance deserves a sceptical reading when you are advising on tool selection. A separate audit, BenchJack, found 219 flaws across ten popular agent benchmarks, with generated exploits scoring near-perfect without doing the tasks. The New Stack

Anthropic's BB+ rating — A Financial Times report that Anthropic was given a BB+ rating — junk, one step below investment grade — by the agency KBRA in June set off a long argument online. Commenters pushed back on the framing: the agency is a minor one, it worked from thin disclosures and produced a risk score rather than a full financial model, and junk ratings are ordinary for a young company. Clients will read the headline, not the thread, so it is worth knowing the counter-argument before someone asks whether their agent platform's vendor is financially shaky. The thread's sharper worry is the lobbying: an investment-grade rating would open the corporate bond market, putting pension and insurance money behind these firms ahead of an IPO investors expect to value Anthropic above $2 trillion. Hacker News