☕️ Microsoft caps engineers' AI spending

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

💸 Microsoft caps engineers' AI spending

💳 Cloudflare gives AI agents a wallet

📌 AWS launches Kiro Crew for 24/7 coding

🔀 GitHub launches stacked pull requests

🤖 Cloudflare open-sources AI triage tool

Plus: 🎁 12 other news you might like, 🧰 6 tools, and 📚 5 papers.

💸 Microsoft caps engineers' AI spending LINK

Microsoft has capped how much engineers can spend on internal AI tools, setting per-division token budget targets as of July 2026 and letting employees track individual spending, while telling staff that maximizing AI use isn't the goal.

To squeeze more value from token spend, Microsoft made OpenAI's cheaper GPT-5.6 the default model for internal GitHub Copilot use, and warned that further restrictions may follow as it monitors consumption.

The guidelines note many engineers currently burn hundreds to a few thousand dollars monthly in tokens; the clampdown mirrors similar throttling at Amazon, Adobe, Atlassian, and Citi despite Microsoft posting rising revenue and profit.

💳 Cloudflare gives AI agents a wallet LINK

Cloudflare rolled out a set of primitives that let AI agents own the full software lifecycle, buying domains, spinning up temporary accounts, and calling the entire Cloudflare API, rather than just generating code.

The release includes `@cloudflare/ci`, a self-healing CI/CD system built on Cloudflare Workflows that can spawn agents, plus OpenTelemetry traces in local dev via Wrangler and the Vite plugin for agent-level observability.

Cloudflare frames this as an "Agent Development Lifecycle," using Workflows and Artifacts to orchestrate dynamic steps, spawning containers, browsers, and agents, setting feature flags, and gradually rolling out changes, so pipelines go beyond linear GitHub Actions YAML.

📌 AWS launches Kiro Crew for 24/7 coding LINK

AWS launched Kiro Crew, an autonomous workspace that keeps AI coding agents running around the clock, letting developers kick off a multi-step task, walk away, and return only when something needs review or approval.

Built on Kiro, Crew orchestrates multiple specialist agents and sub-agents, maintains memory across sessions, runs scheduled jobs, reads existing .kiro config, and carries forward corrections and preferences it learns from repeated developer guidance.

It runs locally or remotely with full audit control, connects via Slack, Telegram and Discord, imports skills from harnesses like OpenClaw and Hermes, and ships as a macOS app on GitHub with Linux and Windows setup guides.

🔀 GitHub launches stacked pull requests LINK

GitHub now natively supports stacked pull requests, letting you break a large change into an ordered series of smaller, focused PRs that each carry context of their place in the sequence rather than being reviewed in isolation.

The feature adds a stack map in the PR UI, stack-wide rule and CI enforcement, and one-click cascading rebases, replacing the manual rebasing and external tooling developers previously relied on for this workflow.

You can review each layer in parallel, then merge one, some, or all at once, with existing reviews, status checks, branch protections, and merge requirements all working without extra setup.

🤖 Cloudflare open-sources AI triage tool LINK

Cloudflare released Flue, an open, platform-agnostic framework for building automated AI agent pipelines, alongside triagebot-action, the GitHub Action that has run issue triage on the Astro repository for several months.

The pipeline reads incoming bug reports, reproduces them in sandboxes, diagnoses root causes, ships preview releases via pkg.pr.new for reporters to verify, then opens a linked pull request, cutting open issues from over 200 to about 30.

It runs as a label-driven state machine in GitHub Actions using isolated subagents per phase-reproduce, diagnose, verify, fix-passing findings through a report.md file, with agent failures treated as signals of missing tests, docs, or unclear abstractions.

Other news you might like

• Why An LLM’s Memory Gets Expensive and How to Fix ItLINK

• Comparing AI agent sandbox platforms: E2B, Modal, Daytona, and moreLINK

• How to Evaluate Voice Agents with LangSmithLINK

• Nvidia’s NOOA makes an agent one Python classLINK

• Deploy local agents everywhere with LFM2.5-2.6BLINK

• Nvidia open sources cuFile API, accelerating GPU read/write capability for high-speed storageLINK

• Can MS Paint run Doom? (Yes)LINK

• API keys vs. OAuth: How to manage both without building two systemsLINK

• How to build an MCP app on the 2026-07-28 spec with WorkOS AuthKitLINK

• New release of LLM adds support for reasoning traces, OpenAI Responses, server-side tools, and smarter loggingLINK

• Your agent can now debug Workers with local tracingLINK

🧰 Trending tools

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

Sim: a workspace for building and deploying AI agents visually or with code, connecting to 1,000+ integrations and every major LLM provider.LINK

ZooData: converts webpages into structured JSON, cutting LLM token usage by 75% while delivering pre-analyzed Amazon and TikTok e-commerce insights for AI agents.LINK

CartAI: automates checkout across merchant sites via a single API call, handling payments, shipping, and order tracking without custom per-site integration work.LINK

ditto.site: turns any public URL into clean Next.js or Vite code deterministically, preserving components, design tokens, and hover states via a free API or MCP server.LINK

Replay QA: records website sessions with devtools access, letting you create detailed, shareable bug reports without running manual QA processes.LINK

📚 Trending papers & reports

Pinpointing slow-code culprits in large software finds the exact spot causing performance drag by comparing runs at different workloads, lifting accurate fix-location detection by ~4 percentage points over existing tools.LINK

Custom chip instructions can now be generated automatically from plain English descriptions, with an AI system correctly writing and verifying these RISC-V processor extensions 73% of the time versus almost never before, cutting the slow manual engineering work chipmakers face when customizing processors for new products.LINK

Chart-making assistants get more reliable by generating several possible interpretations of a vague request, testing each resulting chart, then combining the feedback to pick the best final visualization, beating prior AI methods on standard benchmarks.LINK

Automated requirements extraction using chained chatbots to pull business goals from software documents hit only 61% accuracy, showing it should speed up human analysts rather than replace them.LINK

Automated proof cleanup trims AI-generated math proofs by over 70% on competition problems and over 20% on research code, cutting compilation time up to 60% without retraining the underlying model for each new software version.LINK

See you tomorrow for a new dose of ☕️ Devshot!