Agents, models and security

Controlled coding agents, local model performance and an urgent PostgreSQL patch lead this edition, with judgement remaining the key analytical weakness.

Briefs

Agent engineering

Agents stumble on analytical judgement — Hex’s DataBench technical write-up finds agents handle evidence retrieval better than ambiguous decisions and misleading conclusions. Teams using analytics agents still need human review for open-ended recommendations.
Caveman cuts agent context overhead — The Caveman repository uses recoverable compression for tool output and other agent context. Its benchmark reports 33.2% fewer input tokens across Claude Code runs.
Cheaper development elevates product judgement — A technical write-up argues that cheaper implementation makes deciding what behaviour to encourage more important. Product teams can treat successful features elsewhere as hypotheses, rather than copying them wholesale.

Models and infrastructure

Qwen promotes fast local multimodal model — Qwen’s release announcement introduces Qwen3.8-27B, an open-weight multimodal model with a 262,000-token context window. A performance demonstration claims 65 tokens a second on one RTX 4090.
DeepSeek’s low-cost coding trade-off — A coding comparison and cost and timing report say DeepSeek Pro V4 Max completed a Rust rewrite for roughly $23 in token costs. The run took two and a half hours, making it better suited to batch work than urgent tasks.
AI efficiency chart lacks methodology — A chart post attributes an eighteen-fold rise in AI intelligence per joule over 16 months to hardware and model gains. The figure points to lower operating costs for capable systems.

Policy and governance

Anthropic sets out EU watermarking plans — Anthropic says in its announcement that it is implementing output watermarking for EU AI Act compliance. It says other major model developers that signed the Code of Practice will also implement watermarking.
Amodei questions concentrated regulatory control — Dario Amodei’s statement frames AI regulation around who ultimately controls advanced systems, compute and deployment. The question sits at the centre of the argument over rules that reduce harm without locking power into a few hands.
Amodei defends Anthropic’s risk messaging — In a separate statement, Amodei says his public discussion of AI has given comparable weight to benefits and risks. The dispute concerns how leaders’ warnings shape public expectations and regulation.

Security

PostgreSQL release fixes 28 CVEs — The latest scheduled PostgreSQL release fixed 28 CVEs, including 17 scored above 8.0, according to a security release timeline. Database operators should review the relevant minor releases and patch affected installations.

Tutorials & Demos

A controlled path from specification to pull request

Danny Postma’s system demonstration video shows an agent workflow that moves software work from specification through planning, review, implementation, testing and documentation. Specialised agents work in isolated containers with restricted repository, filesystem and MCP access. An inbox stops the workflow when a person must decide or unblock it. The design also supports recurring jobs and customer-support tasks. The most useful lesson is controlled delegation: give agents narrow access, require approval at important stages, and set firm runtime and spending limits. Postma describes the system’s cost controls after an expensive unattended run.

Worth watching

02:00 — Postma explains how each task runs in an isolated container with restricted repository and filesystem access.
10:04 — He demonstrates a multi-stage feature workflow with specification, planning, review, implementation and approval gates.
21:25 — He describes the system’s cost controls after a goal consumed about $1,000 and reports spending roughly $500 a day.