DoorDash runs 130K tasks on cloud agents

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🤖 DoorDash runs 130K tasks on cloud agents

⚡ AWS launches Graviton5 instances

🐘 Databricks adds Lakebase Postgres autoscaling

🔌 MCP still hasn't solved agent tooling

🦀 rustup 1.29.1 released

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

🤖 DoorDash runs 130K tasks on cloud agents LINK

DoorDash shifted its engineering agents off laptops onto Flux, a cloud platform that ran 130,000 automated tasks in a single month this year, including over 25,000 code reviews weekly across 300-plus reusable playbooks.

Flux isolates each agent in a Firecracker microVM provisioned with repos, tools, and secrets, hitting a p95 under five seconds for full sandbox setup, while an in-house MCP gateway scopes permissions and logs activity for auditing.

Playbooks written in YAML define the task, tools, permissions, and safety boundaries, mixing agent steps with deterministic code, and workflows launch from Slack, GitHub, cron, the CLI, or conversational interfaces after starting with automated code review.

⚡ AWS launches Graviton5 instances LINK

AWS has launched memory-optimized R9g and R9gd instances built on its Graviton5 CPUs, targeting in-memory databases like Redis OSS, Memcached and Valkey, big data analytics, and containerized microservices on EKS.

Graviton5 delivers up to 25% better compute, 35% faster web and ML inference, and 30% faster database processing, backed by DDR5 8800 memory, a five-times-larger L3 cache, and triple the packet-processing throughput of Graviton4.

The instances scale to 192 vCPUs and 1,536 GiB memory across 11 sizes, with R9gd adding up to 11.4TB of local NVMe; a new feature reallocates bandwidth by up to 25% between EBS and VPC networking.

🐘 Databricks adds Lakebase Postgres autoscaling LINK

Databricks has rolled out autoscaling for Lakebase Postgres, dropping upfront instance sizing by resizing VMs in place based on live workload signals, made possible by an architecture that fully separates compute from durable storage.

The autoscaler picks the largest target from three signals, CPU load kept at or below 90%, memory held under 75% of RAM (polled every 100ms to avoid kernel kills), and a working-set estimate that sizes the local file cache to fit hot pages.

Working-set sizing uses a time-windowed HyperLogLog storing timestamps per register to estimate distinct pages accessed over one to 60 minutes, while NeonVM applies CPU and memory changes to a running VM and can live-migrate it without dropping connections.

🔌 MCP still hasn't solved agent tooling LINK

A new open spec called Agentic Resource Discovery (ARD) targets a gap MCP left open: it lets agents search for the right tool across many registries instead of assuming the client already knows which server to call.

Released under Apache 2.0 and authored by engineers from Google, Microsoft, and Hugging Face, ARD's v0.91 proposal uses JSON-LD and a REST interface with a required POST /search endpoint that searches by task, plus optional browse endpoints.

ARD federates existing catalogs without forced migration, letting companies keep internal resources private while routing searches to approved external services; AWS is exploring wiring it into its Agent Registry, and governance may move to a neutral body like W3C.

🦀 rustup 1.29.1 released LINK

The rustup team shipped version 1.29.1, adding faster concurrent operations, a new local docs server, official aarch64-pc-windows-gnullvm host support, and a deprecation warning for implicit toolchain installation in several commands.

Concurrency now speeds two paths: `rustup update` checks for available updates in parallel, and `rustup component add` installs multiple components at once, while `rustup doc --serve` exposes docs over local HTTP for containerized setups.

Installing i686-pc-windows- toolchains on 64-bit Windows now requires `--force-non-host`, "target triple" was renamed to "target tuple" project-wide without breaking the CLI, and cancelled or failed installs no longer leave stray files.

Other news you might like

• Introducing wraptureLINK

• Testing cookie behavior across hundreds of web surfaces with our in-house auditorLINK

• How Enterprising Modders Installed Windows NT On Nintendo GameCubesLINK

• Kubernetes v1.37: Storage Version Migration Enabled by DefaultLINK

• Linux 7.3 Features Many Exciting Improvements, New Hardware Support & Faster BtrfsLINK

• Git Submodules as a Package ManagerLINK

🧰 Trending tools

Sim: build and deploy AI agents through a visual or code workspace, with 1,000+ integrations and support for all major LLM providers.LINK

Prefactor: an evaluation layer that scores agent runs in real time, catching quality regressions and drift before they reach production users.LINK

Dograh: open-source voice agent platform with a visual flow builder, telephony, human transfer, and self-hosted QA monitoring across 30+ model integrationsLINK

Zro: routes coding requests to open-source models like GLM-5.2, DeepSeek V4 Flash, and Kimi K3 through one endpoint without storing your dataLINK

Meridian: open-source activity tracker running locally on your device, logging work with context so you can see exactly where your time wentLINK

AgentSky: runs managed AI coding agents with auto-recovery, controllable from WhatsApp, Telegram, Slack, web, API, or CLI for remote developmentLINK

📚 Trending papers & reports

Protein-folding data storage shrinks the bulky output from large-scale AlphaFold 3 experiments by roughly 5 to 15 times, giving drug and protein-design teams faster, cheaper access to their prediction results and confidence scores.LINK

Silent trial software for pathology AI lets hospitals test diagnostic models on live patient slides without affecting care, providing a reusable system already running at three institutions to evaluate cancer mutation prediction from tissue images.LINK

Chip design assistants get a system that lets AI break complex hardware into blueprints and test them automatically, correctly producing working circuit designs ~96% of the time, beating prior AI tools.LINK

Rust's safety guarantee has holes, with a study of 30 real bugs showing the official compiler sometimes accepts programs that break memory safety, flaws that often sit undetected for years after a feature ships.LINK

Climate-science Q&A improves when an AI search tool maps how concepts connect across many articles instead of reading each in isolation, boosting answer relevance by 160% and completeness by 177% over standard retrieval.LINK

See you tomorrow for a new dose of ☕️ Devshot!