☕️ The blank-check AI coding era is over

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

💸 The blank-check AI coding era is over

🧠 Liquid AI runs agents on Raspberry Pi

🐕 AWS's Dogwood fixes wrong AI tool calls

🌳 Anthropic wants one git worktree per agent

🎮 Kubernetes DRA overhauls GPU scheduling

🌐 Cloudflare turns any site into WebMCP

Plus: 🎁 5 other news you might like, 🧰 6 tools, and 📚 5 papers.

💸 The blank-check AI coding era is over LINK

Microsoft has ended unrestricted AI coding spend, rolling out per-division token budgets since July, giving engineers dashboards to track their own usage, and setting OpenAI's GPT-5.6 Sol as the default model in GitHub Copilot.

Sol remains OpenAI's priciest GPT-5.6 model at $5 per million input and $30 per million output tokens, versus Terra at $2/$12 and Luna at $0.20/$1.20, pushing teams to weigh capability against cost per task.

Setting a default gives Microsoft control over where token spend goes after Copilot's auto-routing sometimes sent engineers to Anthropic models, while GitHub's June 1 usage-based billing added org-level and per-user budget controls.

🧠 Liquid AI runs agents on Raspberry Pi LINK

Liquid's new LFM2.5-2.6B open-weight model runs agentic workloads entirely on local hardware-from smartphones and laptops down to a Raspberry Pi-with no cloud inference or GPU required, targeting edge and privacy-sensitive deployments.

The 2.6B-parameter model packs a 128K-token context and native tool calling, decoding ~220 tokens/sec on an Apple M5 Max and ~30 on a phone in under 2.5GB of memory, with day-one llama.cpp, MLX, vLLM, SGLang, and ONNX support.

Trained inside real agent harnesses like Hermes Agent and OpenClaw, it tops instruction-following and most tool-use benchmarks against larger Gemma 4 and Qwen3.5 models, but ships under a license requiring companies above $10M revenue to strike a commercial deal.

🐕 AWS's Dogwood fixes wrong AI tool calls LINK

AWS released Dogwood, an open-source policy language and reference interpreter under Apache 2.0 that governs entire sequences of AI agent tool calls rather than evaluating each action in isolation, extending its Cedar authorization language.

Dogwood adds temporal conditions built on Metric First-Order Temporal Logic, letting policies check prior events, count calls in a time window, or sum transferred amounts-so a stock-selling agent, for instance, can act only after a matching approval within the past hour.

Because it tracks in-flight requests, Dogwood catches parallel tool calls that would evade completed-transfer limits, but it's stateful and pricier than Cedar, and the reference interpreter targets testing rather than production authorization.

🌳 Anthropic wants one git worktree per agent LINK

Anthropic's docs now recommend running each parallel coding agent in its own git worktree by default, turning what was once an expert workflow into the standard starting point for supervising multiple agents at once.

A worktree hands every agent a private repository copy per checkout, letting ten work without seeing each other's edits, but Faros AI telemetry across 10,000+ developers shows high-AI teams merge 98% more PRs while review time grows 91%.

The real bottleneck sits below code: four agents' changes queue behind one staging cluster, one seeded database, and one message queue, and blocked agents either idle on stale views or validate against mocks, forcing rework once the shared environment frees up.

🎮 Kubernetes DRA overhauls GPU scheduling LINK

Kubernetes 1.34 shipped Dynamic Resource Allocation, replacing the crude `nvidia.com/gpu: 1` integer counter with a scheduling model where GPU drivers publish structured hardware data that workloads can query by memory, generation, and topology.

Using Common Expression Language in a ResourceClaimTemplate, a job can now request "an H100 or better with at least 40GB", fall back across MIG slice sizes to a full GPU, or demand four NVLink-connected cards.

The approach collapses per-GPU-type node pools, taints, and duplicated Helm charts into a single manifest, ending the pattern where jobs OOM on undersized cards while larger idle GPUs sit unused nearby.

🌐 Cloudflare turns any site into WebMCP LINK

Cloudflare launched a developer preview that lets any site it hosts expose WebMCP tools to browser agents with a single dashboard toggle, requiring no code changes and nothing altered at your origin.

WebMCP is a new browser standard shipping experimentally in Chrome 146 as document.modelContext, letting sites register tools agents can call rather than guessing through pages built for humans, saving tokens on tasks instead of navigation.

Cloudflare uses HTMLRewriter to inject a bridge script that composes tool packs and registers each via .registerTool; the preview ships Content Credentials and Site MCP Server packs, both running entirely in the visitor's browser with no server round trip.

Other news you might like

• Deep Agents vs LangChain vs LangGraphLINK

• 2026 State of CSS, Devs SurveysLINK

• Humans in the loop miss a third of dangerous AI coding agent requestsLINK

• Qwen 3.8-Max and Claude Opus 5 show why raw benchmark scores don't predict the billLINK

• Free agents: How AWS Kiro could untie agents from editorsLINK

🧰 Trending tools

AnySearch: a search API for AI agents that pulls filtered, de-duplicated, structured results from trusted sources in parallel, improving reliability.LINK

SKI: lets you voice-code with Claude Code, Codex, and other agents, getting spoken replies back so you build hands-free at thinking speed.LINK

Humalike x Hermes: provides turn-taking, timing, and memory APIs so AI agents know when to speak, wait, or interrupt naturally in conversation.LINK

Openbase: helps developers pick reliable open-source packages by comparing popularity, activity, and reliability metrics alongside real user reviews.LINK

Pushary: sends AI coding agent approval requests to your phone's lock screen, letting you approve actions remotely so Claude Code or Cursor keeps working uninterrupted.LINK

Claude Code usage tracking by LangWatch: monitors token usage, costs, and performance metrics for Claude Code sessions, helping teams debug and optimize AI agent behavior in production.LINK

📚 Trending papers & reports

File-finding assistant for AI coding agents cuts the searching busywork before a bug fix, resolving 27.0% of test issues versus 25.8% without it, using 15% fewer steps and 19% fewer tokens.LINK

Coding agent skill-building lets AI dev agents update their playbook of reusable techniques instead of relearning skills per task, lifting bug-detection and false-positive-filtering accuracy, with F1-score up ~61 points in a real industrial deployment.LINK

Diagram-generating chatbots now have a standardized 132-sample test that scores whether they can turn plain-English descriptions into correct software sequence diagrams, exposing large, previously unmeasured gaps between models.LINK

AI bug-fixing agents that autonomously explore code, plan fixes, and act on feedback now have their progress mapped across 242 studies, showing reinforcement learning as the key method pushing automated software repair forward.LINK

Coding agent instructions can be auto-tuned to balance accuracy and token spend, cutting costs up to ~32% while boosting task success by up to 21 percentage points versus fixed, hand-written skill sets.LINK

See you tomorrow for a new dose of ☕️ Devshot!