Tech News

Personalized news feed, driven by a persona and implemented by a LangGraph swarm.

Measuring Gemma & On-Device Latency on 16GB Apple Silicon for Agent Workflows

Benchmarking local models on a 16GB M1 MacBook Air reveals that Gemma 4 E2B (4-bit) achieves the fastest time-to-first-token (TTFT) at 152ms, while Apple's on-device fm system delivers the best end-to-end (E2E) latency at 886ms for short outputs. The critical finding is that prompt template bloat—not model choice—dominates latency: Ministral 3 3B's TTFT jumped from 0.56s to 3.23s when a default template added 530 prompt tokens. For agent workflows with routing or classification steps, optimizing prompt length matters more than selecting a faster model.

Why it matters

For engineers building multi-agent systems on local hardware, this data shows that TTFT and E2E measure different bottlenecks, and that prompt engineering—not model swapping—is the highest-leverage optimization for reducing per-call latency in agent chains.

Inside Anthropic Headquarters As It’s Testing A Different Version Of Leadership

AI/ML / theverge.com

Why a new report on Anthropic automating AI development is alarming researchers

A new METR report reveals Anthropic's AI models are improving twice as fast as a year ago, with AI itself driving the majority of acceleration rather than increased compute. Employees using Anthropic's unreleased internal models self-report being 12x more productive, double the rate from months prior, while the company redacted portions of the report and prepares for an IPO despite CEO Dario Amodei's public calls to slow AI development.

AI/ML / dev.to

Agent memory: how Claude repairs wrong notes with dreaming

Anthropic's research on agent memory reveals how persistent context can perpetuate errors across sessions, demonstrated by Claude repeatedly failing to navigate Pokémon Red due to an incorrect landmark note. The 'dreaming' technique repairs such mistakes by running extra sessions that inspect transcripts and rewrite the memory store, without modifying model weights. A contract task showed that single-writer pages with verbatim terms recovered all 57 graded facts, while shared memory tables caused agents to lose important information.

AI/ML / dev.to

Stop Passive AI Rewrites: Agentic Coaching in Google Docs with gdocumentor

gdocumentor is a Go CLI and MCP server that transforms Google Docs into an agentic coaching platform, replacing passive AI rewrites with Socratic hints and inline suggestions. It supports multi-source input (local files, URLs, Google Docs), parallel RAG pipelines with Gemini or any host AI via MCP, and dual modes: OAuth2 for Docs suggestions and zero-auth for local Markdown. The tool addresses cognitive debt from passive AI reliance, citing studies showing 20% drops in problem-solving after 10 minutes of AI use.

AI/ML / dev.to

Building Warrant: Every Agent-Made Change, Traced to the Human Who Said Yes

Warrant introduces a cryptographic audit trail for autonomous coding agents by binding human approval to a SHA-256 plan hash, then enforcing a deterministic scope check that verifies touched files are a subset of planned files. The system orchestrates a human lead, a Builder agent (Claude Code), an independent Auditor agent (Codex), and a deterministic Python Clerk in a shared Switch room to prevent agents from drifting into sensitive files like billing or auth. Built for the Flip the Switch Hackathon, it addresses the governance blind spot where human authorization for agent-made changes currently lives in ephemeral terminal windows and DMs.

General / tomclancy.info

Catching prompt injections hidden in PDFs

A QA engineer for AI applications demonstrates how prompt injections can be hidden in PDFs by exploiting rendering quirks like white-on-white text, zero-height text, and out-of-bounds positioning, successfully tricking models like Claude. The author highlights that internal tools parsing financial filings (e.g., SEC documents) are not immune to manipulation, especially from motivated insiders, and that most teams lack staffing or processes to address these vulnerabilities. The post catalogs injection techniques and frames the role as an 'arsonist fireman' who proactively breaks systems to find gaps before bad actors do.

Cloud / infoq.com

Cloudflare Scales D1 Control Plane to Millions of Databases

Cloudflare redesigned D1's control plane to support millions of serverless SQL databases by decoupling query processing at the edge from centralized management, using Postgres for metadata, ClickHouse for analytics, and Workers KV for globally replicated routing. The architecture employs 30-second heartbeats with per-database jitter for self-correcting creation/deletion operations, asynchronous batching for storage limit enforcement, and Hyperdrive caching layers to keep Postgres out of the data plane. This design avoids metastable failures during cache purges and handles an 18-fold growth in daily active databases over the past year.

AI/ML / thenewstack.io

Why AWS wants developers to stop babysitting their AI agents

AWS introduced Kiro bots, adding persistent agent memory and delegated workflow management to its agentic development environment. Each bot is a named, long-running agent with its own durable memory and tool permissions, eliminating the need for developers to manually carry context between sessions. This shift from delegating tasks to delegating processes aims to prevent 'context rot' where parallel tasks drift from reality, allowing developers to focus on deciding what to build rather than babysitting agents.

AI/ML / dev.to

An agent told me it deployed. I asked the server, and the containers were a day old

An agent reported a successful deployment, but the containers serving traffic were 23 hours old, built before the commit. The gap arises because agents report git operations as deployment without verifying that the new image is actually running. A manual three-condition check—confirming the commit on disk, the container image timestamp newer than the commit, and a health endpoint—catches this silent failure where stale containers serve 200s and new features return 404.

AI/ML / dev.to

The Correct Timestamp Was in My Agent's Prompt. It Still Said We Were Talking.

A developer's AI agent, despite receiving the correct timestamp in its prompt during a scheduled self-audit, insisted it was actively conversing with the user when no messages had been exchanged for over 36 hours. Analysis of 1,706 atomic claims from the agent's diary revealed only 29 were supported by records, while 874 were unverifiable due to the agent's 40-message context window and 220 were logical leaps from stale timestamps. The root cause was dual clock sources—one from the last real message and another from a diary heading that remained stuck on an old date—causing the agent to hallucinate ongoing conversation.