Writing

Your prompt is cut into tokens, each token becomes a vector of 1,024 numbers, 28 transformer layers rework those vectors, and out comes a probability for every possible next token. Pick one, add it on, repeat. I measured each step on a small open model: cat sits nearest kitten, 42 of 448 attention heads link "it" to "animal", and it confidently named the capital of a country that doesn't exist.

After fixing the problems I found, WebMCP Resilience caught 15 of 19 timing bugs, up from 11, with no false alarms on correct code. Each fix earned its place, and one introduced a bug of its own that the rerun caught.

An ordinary check caught none of 19 timing bugs. WebMCP Resilience caught 11, with no false alarms on correct code, and showed me exactly where it falls short.

How I keep experimenting with GenAI without spending a fortune.

Same shop, same tools, four architectures, 4,480 runs. Simple and branching emails were handled at or near 100% without an agent. One model call coped with unexpected requests too. The agent won when something failed mid-way (74% against 42–50%), though two of those cases are simple fixes I left out of the fixed code. It cost about the same as the others, and with a cheap model it refunded a timed-out refund twice in 16 of 20 runs.

A private key instead of a client secret closed the stolen-credential takeover; a confirmation the service insists on, given on the approver's phone, stopped the £4,800 invoice every time; agent capability checks, claimant details and a locked tool server closed three more gaps. A £900 claim with no receipt and a fake pre-approval was still approved in 4 of 5 runs.

Four common ways to give an AI agent access to company systems, tested with a local Keycloak and planted requests. A broad service account let 11 of 20 planted requests through; delegated, task-scoped tokens with a confirmation step let none through, cut off a leaver in 110 seconds and shrugged off stolen tokens. It was the combination of controls that worked, none of them checked whether a claim was genuine, and agent hand-offs broke the chain.

Coding agents leave a one-line note when a repo slows them down, git hooks record it, and a weekly report shows the gaps worth fixing.

Notes from building with WebMCP: humans and agents sharing the same browser state.

Claude Code 2.1.314 adds a setting that blocks writes outside your working directories, the Claude Platform gets a beta usage and cost analytics API and a 128 MB request limit, and Codex CLI 0.166.0 adds GPT-6.1 Terra to its Vertex AI and Amazon Bedrock catalogs.

ChatGPT Pro and Business subscriptions now include monthly OpenAI API credits, Grok Bot 0.78.1 adds shopping and custom-mode cloud agents, LiteRT in Google Play services gets faster GPU inference, Gemini CLI v0.66.0 is mostly fixes, and Anthropic opens a beta Claude Tag API for Enterprise organisations.

Google's Gemini API CLI moves to beta, Codex CLI 0.165.0 adds a built-in Vertex AI provider, Claude Code 2.1.312 changes the system prompt for custom cloud deployment names, and Grok Bot 0.77.0 connects to 1Password without a shared vault.

OpenAI releases GPT-6.1 Terra for computer use at half the API price of GPT-5.6 Terra, the Cursor Agent SDK is generally available, Claude Code 2.1.311 fixes long and unattended sessions, new Gemini API projects store requests by default, and Grok Bot 0.75.8 tightens Team Bot sharing.

Claude Code 2.1.310 brings the Monitor tool to Bedrock, Vertex and Foundry, Grok Bot adds a Vault for saved logins and cards, Grok Build groups agents by folder, and Gemini in Sheets can analyse BigQuery data in plain language.