Agents favor hackable devices; GLM cuts Revolut coding costs 90%
The issue covers open firmware in the agent era, a minimal Claude Code agent setup, a privacy code review benchmark, Revolut's self-hosted GLM models, and several new agent tools and papers.
Top stories
Geoff Stearns: AI agents will favor devices owners can reflash, pushing makers toward open firmware

- Geoff Stearns argues that as AI makes firmware cheap to write, agents will favor consumer devices whose firmware owners can replace, pushing makers to open up.
- M-Vave open-sourced its FM-1 synth firmware under GPLv3 in October with a development kit and browser emulator; Behringer, Akai and Erica Synths made similar moves.
- He says closed makers resist through measures like Bambu Lab's server authorization and Google's developer verification rule, and that owner-held signing keys separate security from lock-in.
Mathew Mathew builds AI agents as plain folders run headless with Claude Code

- Independent AI consultant Mathew Mathew describes dropping frameworks, vector databases and multi-agent setups in favor of an agent that is a folder with AGENTS.md, skills, context and minimal tools.
- The agent runs unattended with claude -p from cron jobs or webhooks, and he reviews saved session transcripts under ~/.claude/projects to decide which instructions to change.
- He says to solve the basic case first and add machinery later, and to measure improvement with evals across harnesses such as Gemini CLI, Codex or pi.
PrivacyPR dataset of 4,131 GitHub pull requests benchmarks privacy-aware code review

- A paper on arXiv presents PrivacyPR, a dataset of 4,131 merged pull requests from 1,469 repositories, including 1,169 high-confidence privacy fixes and 1,643 non-fix PRs.
- On the 100+100 Small subset, two coding agents caught 42-51% of reintroduced privacy defects under a generic review prompt and 88-91% when asked to check privacy, flagging 11-15% of non-fix PRs.
- The authors say detection depends strongly on review intent and comes with false alarms; data, code and an evaluation harness are public.
In brief
- Pizza open-source AI agent replaces the while-true loop with an event-driven state machine. Pizza stores conversations and tool calls in an immutable SQLite event log, gives the model one CLI tool, supports branching and context compaction, and offers desktop, terminal, JSON-RPC and print interfaces. (pizza.hertzbeat.com)
- Vibor Cipan shows how to run multiple Claude Code accounts on one Mac. Vibor Cipan describes a shell function setting CLAUDE_CONFIG_DIR to a separate folder such as ~/.claude-b, so two accounts stay signed in, with Desktop on one and the CLI on the other; tested only on macOS. (twitter.com)
- Revolut says self-hosted GLM models cut AI coding costs about 90% for 2,000 engineers. BigGo reports Revolut's David Abitbol says GLM on 56 Nvidia B200 GPUs costs 30 cents per million tokens, about one-tenth of prior Anthropic and OpenAI spend, with time-to-first-token cut from over 10 seconds to about 1.6. (finance.biggo.com)
- Althar open-source IDE has a coordinator agent and cross-lab model review. Althar is a free Apache 2.0 desktop IDE for macOS, Windows and Linux where a coordinator assigns tasks to agents like Claude Code, Codex and Copilot, with optional review by a model from another lab. (github.com)
- Claude Opus 5.5 designs a roughly 200-component analog synthesizer in 19 hours. The author used Claude Code to handle circuit design, ngspice simulation, KiCad layout and firmware for a Juno-style monophonic synth; hardware matched SPICE closely and needed only one solder-bridge fix. (cpldcpu.github.io)
- HerAgent block-level checkpointing lifts coding-agent environment setup success to 85.0% on Installamatic-Bench. A paper on arXiv reports HerAgent beat six prior agents on four datasets, raising SetupBench success from 46.3% to 70.4% and cutting setup costs by up to 94.5%. (arxiv.org)
- State-machine review finds 85% of known state bugs, versus 71% for default review. QuickChat.ai says on 34 pull requests with 105 known state bugs, its eight-lens state-machine review in QuickReview found 89 against 75 for its default review, at the same cost. (quickchat.ai)
Compiled, edited and written with AI. Published by Rapid Review, Inc.