Loading…
Latest reads
Every post, summarised
Grounded summaries of engineering blogs, newest first. Each links to the original.
- 4,671 posts
- 20 companies
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Latest posts
Announcing the Databricks Big Book of AgentOps
The Big Book of AgentOps presents AgentOps as an operating discipline for building, deploying and improving production AI agents, combining architecture, evaluation, observability, governance, security and cost management. It addresses the operational risks of tool calls, enterprise-data retrieval, API actions, retries and multi-step execution, where quality, permissions and spend can become difficult to control. The six-chapter book covers four agent architectures, four deployment patterns, a seven-phase lifecycle, DevOps practices, operationalization planning and stakeholder alignment, with Unity Catalog, Unity Gateway and MLflow central to its Databricks-based deployment guidance. Its practices include building evaluation datasets from real traces, calibrating automated judges with subject-matter-expert feedback, designing tracing and access controls into systems, and using RACI matrices to clarify ownership. The announcement cites customer examples including FactSet’s 44% accuracy improvement, DXC Technology’s 30% total-cost-of-ownership reduction, and ICE’s 77% syntactic accuracy and 96% execution matches.
Modal is expanding in Europe with our new London office
Modal is expanding in Europe with a new London office, its second in the region, representing a larger investment in the area. The company already has nearly 20 engineers in Stockholm who built core parts of Modal, including its custom file system and Sandboxes product. Hugh Killingbeck-Jones joins as VP EMEA to lead go-to-market teams in London, which the company considers the AI capital of Europe. European AI-native companies like Black Forest Labs and Legora, as well as global platforms like DoorDash, run on Modal for tasks such as model evaluation and agent testing. The company is hiring across sales, engineering, and other roles in London and Stockholm, and plans meetups across EMEA.
How Botika runs full-stack generative AI on Modal
Botika builds agentic e-commerce teams that automate visual production for global fashion brands, using proprietary foundation models and a 100-terabyte image dataset. The company runs its entire full-stack AI operation on Modal, with the data pipeline processing 100 terabytes at less than 1% error rate across thousands of containers. Before Modal, the team struggled with Kubernetes fleet management and slow GCP Batch jobs; now a single researcher can launch 50 experiments daily using Claude, and they set up reinforcement learning infrastructure in less than a day. Modal handles autoscaling, cold starts, and multi-node training with RDMA, allowing Botika to double traffic in seconds, and they serve about 15 models on L4s, L40S, A100s, and H100s. The infrastructure is not Modal-specific, enabling the team to run open-source and custom frameworks without modification.
Self-hosted machines
Cursor’s self-hosted machines keep tool execution inside an organization’s own network, leaving the codebase, build outputs, and secrets on internal infrastructure while the agent handles tool calls locally. The feature supports My Machines for connecting one laptop or VM and team pools that provide named queues of workers, with capacity expanding as requests arrive and contracting when workers disconnect. Pools are repository-independent, can hibernate idle machines, and can restore them within a reconnect window, avoiding the need to keep costly capacity warm. Cloud agents can run on infrastructure including AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel, and E2B, while Linux and Mac workers add computer use for clicking, typing, screenshots, and browser control.
4 engineering patterns behind the strongest AI Agents Challenge submissions
The strongest submissions in the Google for Startups AI Agents Challenge repeatedly used four engineering patterns rather than merely labeling prompt chains as multi-agent systems. They exposed internally used tools through bidirectional MCP, used typed events and asyncio queues to let independent agents run concurrently, and routed fallback outputs through one shared validation function. They also reduced inference cost with tiered routing: regex or a cheap classifier handled simple intent before requests reached an expensive reasoning model, with one team reporting more than 40 percent handled before a real model call. The post argues these practices improve bounded data access, latency, resilience, and cost without requiring larger teams or newer models, and notes that Agent Development Kit and Agents CLI submissions showed them frequently.
Google for Developers Blog - News about Web, Mobile, AI and Cloud
The page presents six featured Google developer articles spanning multi-agent systems, content credentials, TPU optimization, agent plugins, Go, and Genkit Go. One article identifies bidirectional MCP, async event buses, unified validation, and tiered routing as recurring patterns in successful AI Agents Challenge submissions. Another introduces Credentio, an open-source C++ library for local-first C2PA validation, while a HeyGen case study reports a 1.86x speedup for Avatar IV on Trillium TPUs using torchax, XLA, FSDP, Ulysses sequence parallelism, and custom kernel/compiler optimizations. The collection also covers Agent Plugins 1.0.0’s plugin.json directory specification and Genkit Go’s progressive-disclosure Agent Skills based on SKILL.md bundles. A final article argues that Go’s compiler, toolchain, readability, and backward compatibility provide guardrails for reviewing and maintaining AI-generated software.
How River takes security work from a fix to merge
Shopify’s River is an AI agent in Slack that manages dependency and first-party application vulnerability remediation beyond the patch. It validates ledger entries against repository, pull-request, and vulnerability-tracker state, checks whether findings remain active, and updates code on the repository head when it can do so safely. River regenerates lockfiles with repository tooling, repairs upgrade failures, discards stale CI results after rebases, and avoids overwriting human work. When product or risk judgment is required, it preserves evidence and hands owners a question, then verifies the repository head, dependency graph, tracker, and ledger after merge. In its 11 days, the dependency backlog fell about 70%, while security merges through the freshness-gated queue rose from about 10% to 80%.
From Neon Postgres to ClickHouse Managed Postgres
This user story describes three teams that moved production Postgres workloads from Neon to ClickHouse Managed Postgres after encountering reliability, performance, cost, and network-transfer problems as usage grew. ClickPipes handled schema migration, a parallel initial load, and change data capture, allowing teams to test synchronized databases before cutover. Reported outcomes included zero downtime or connection restarts, faster and more predictable queries, lower CPU or transfer costs, and p99 latency improving from seconds to single-digit milliseconds in one case. The post attributes these gains to local NVMe storage and eliminating transfer fees when analytics already runs on ClickHouse, while noting differences from Neon such as slower branch creation and manual application of several DDL changes.
REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs
REFACTOR-VLA addresses the weakness of monolithic vision-language-action (VLA) models, which generate raw commands or short action sequences and struggle with long-horizon tasks and interpretability. Its wake/sleep architecture uses a Behavioral-Equivalence Kernel (BEK) to cluster motor-program segments by outcomes from a learned latent world model, then produces typed lambda terms for a library-conditioned rectified-flow action decoder. Skills are admitted only when they satisfy both Minimum Description Length (MDL) and return-preservation gates, while training proceeds through world-model warmup, wake-phase policy optimization, and sleep-phase skill discovery. On the full LIBERO benchmark, expanding the world model from 188 million to 430 million parameters worsened results on all four suites, showing that scale alone was not beneficial. Adding supervised contrastive InfoNCE loss during warmup substantially improved sleep-phase clustering, with reported NMI scores of 0.462, 0.867, 0.915, and 0.754 for the Object, Spatial, Goal, and LIBERO-10 suites, respectively, under n = 3 multi-seeding.
Shared Nothing Architecture
Shared-nothing architecture is presented as a scaling model for Showcase, a Ruby on Rails ballroom-competition application that began on a Mac Mini and grew to cloud VMs serving events in multiple regions. Because each event has few geographically concentrated users, a roughly one-megabyte SQLite database, about one transaction per second, and typical response times of 100ms or less, the design keeps state separate rather than sharing databases and Redis across events. The implementation runs an event’s services on one Fly.io VM, repeats that unit across machines or regions, and can place multiple event instances on one server with Phusion Passenger, NGINX, volumes, and a shared Action Cable process. It specifies 1GB machines with 2G swap after observing about 220MB idle and 450MB active RAM, and recommends WireGuard; the conclusion is that local state and fewer network calls suit this workload and can improve throughput and reliability.
How the FDA is building a secure, AI-ready data foundation on Databricks for Government
The FDA built HALO (Harmonized AI and Lifecycle Operations for Data) as a secure, governed, AI-ready enterprise data platform for modernizing siloed systems without interrupting regulatory work. Its move to Databricks on AWS GovCloud, following FedRAMP High authorization sponsorship, added Unity Catalog as a governance layer across a multi-tenant architecture, while Terraform-based security patterns, PrivateLink, customer-managed keys, and the compliance security profile support regulated workloads. The agency migrated more than 5,000 users and 8,000 jobs and pipelines with zero downtime, refactoring over 1,000 pipelines and 4,000 notebooks. After onboarding eight centers and 30 programs, FDA reported query responses improving over 30%, compute costs falling over 20%, and provisioning and sharing time dropping over 75%. HALO also supports responsible AI use cases such as MARS, with humans retaining decision authority.
BenchMIRT: What are LLM benchmarks actually measuring?
BenchMIRT is a method for auditing LLM benchmarks at the individual-prompt level, where questions may measure capabilities beyond a benchmark’s stated goal. It applies multidimensional Item Response Theory (MIRT) to model and question results, estimating model abilities plus each item’s difficulty and discrimination across dimensions. Trained on 100 LLMs, 16 benchmarks, and more than 34K questions, it independently recovered safety and general reasoning as dominant dimensions, repeatedly producing the same result. The analysis found, for example, that BBQ and WMDP aligned more strongly with reasoning than safety, while HarmBench combined safety- and reasoning-associated question groups; retaining 10% of items generally preserved capability rankings, and held-out correctness prediction reached 79% versus 70% for a baseline. BenchMIRT’s dimensions depend on the benchmark set, and its item-level transparency can also enable evaluation trimming.
How we eliminated $1 million a year of wasted AI agent spend in one hour
Databricks engineers investigated whether silent failures in MCP tool calls were driving hidden AI-agent costs as usage increased. Unity Gateway emitted OpenTelemetry traces for every invocation, recording tool names, arguments, errors, token counts, latency, and session IDs in a table that Genie One could query with plain-English questions. In a 24-hour sample, the analysis identified seven Jira and Google Drive/Docs bugs responsible for an estimated $499,000 annually in wasted tokens and 12,023 hours of agent wait time, or $1.2 million in lost productivity. The most costly patterns included under-specified interfaces, incompatible argument shapes, and cryptic or misleading errors that triggered retries and recovery loops. Engineers fixed the issues with a coding agent, completing the find-quantify-fix cycle in about an hour; the post recommends designing tools to tolerate reasonable LLM inputs and continuously tracing failures.
ClickHouse welcomes RunReveal
ClickHouse has acquired RunReveal, a company that built a security data platform on ClickHouse, and says the team’s expertise will become part of its work in security. The announcement frames security as a demanding data workload involving cloud audit logs, identity events, endpoint telemetry, and network flows, with continuous high-throughput ingest, years-long retention, and interactive query requirements. ClickHouse says the acquisition will bring sharper reference architectures, guidance on schema and retention design, and a clearer connection between customer requirements and its database roadmap, while RunReveal remains available through a bring-your-own-database model. Existing RunReveal customers will continue to receive platform support under unchanged contract terms, and the team’s work on agentic investigation is expected to inform support for analytics involving agents that hunt across sources, tune detections, and advance investigations.
The Agentic Analytics Benchmark: Measuring model accuracy and efficiency in analytical agents
The post presents data-agent-mnist, an open harness for measuring analytical agents on correctness, cost, and speed against a warehouse-shaped workload rather than a single text-to-SQL query. The authors extracted 201 production analytics questions from DWAINE, curated and anonymized them, reconstructed a deterministic synthetic warehouse with 18 tables and 865 columns, and evaluated 28 proprietary and open-weight models. The benchmark preserves schema discovery, multi-turn querying, planning, and multi-hop joins across curated marts and a dimensional CRM layer, with provider-diverse ground truth and judging described as part of the workflow. On the authors’ warehouse, Claude Fable 5.1 ranked first for correctness at 76.6%, while the full run cost $1 with DeepSeek V4 Flash versus $52 with Fable 5.1, an 11-percentage-point correctness difference. The harness is intended to identify the best production fit for a specific warehouse, not a universal winner.
Collaboration makes us all stronger
An external researcher, Mehmet Ince, reported a memory-safety flaw in PostGIS’s address_standardizer extension, where caller-controlled grammar data indexed a fixed-size array without bounds checking. Because the extension was available to ordinary tenant roles on managed Postgres services, Databricks treated the issue as a platform exposure rather than a third-party problem, while noting that its microVM architecture meant the exploit did not produce cross-customer impact on Lakebase Postgres and Neon. Detection alerted the Neon team during testing, enabling rapid validation and downstream patching while a durable fix was developed and deployed to protect tenants without customer action. The upstream fix initially missed cases; Mehmet supplied the remaining pieces, and the complete correction was returned to PostGIS, with his bounty donated and matched for the project.
How Claude's text watermarking works
Anthropic describes a text-watermarking system planned for Claude to estimate whether Claude was involved in generating or processing text, primarily to comply with the EU AI Act. The method uses a key and preceding words to influence the source of randomness in low-stakes word choices, creating a detectable pattern without adding characters, tokens, identifying information, or extra cost. Anthropic says internal testing found no impact on content, creativity, readability, or perceived quality, while detection confidence improves with longer passages and weakens when text offers few valid alternatives, such as factual statements, proofreading, or code. The watermark can indicate only that Claude was likely involved, not whether text was human-written, who authored it, or which person or organization used Claude; light editing may not remove it, but a complete rewrite will.
New system views in PostgreSQL 19
PostgreSQL 19 adds four system views for lock contention, recovery, autovacuum priorities, and dynamic shared memory allocations. The article presents pg_stat_lock as cumulative, cluster-wide statistics with one row per lock type, reporting waits, total wait time, fastpath_exceeded, and stats_reset; waits and wait_time count only successfully acquired locks that exceeded deadlock_timeout, while fast-path overflow attempts are counted separately. Examples reset statistics, reproduce a 2.5-second table lock, and scan 140 partitions, producing 1 wait, 2201.783 milliseconds, and 422 fast-path overflows. The post describes recovery and autovacuum views, noting that autovacuum scores suggest likely priorities rather than guarantees because scoring and worker scheduling can differ. DSM registry entries expose name, type, and size; NULL means initialization failed.
Building an Adaptive Agentic Cybersecurity System with NVIDIA Nemotron
NVIDIA and CrowdStrike describe an agentic cybersecurity system that links red-agent attack execution with blue-agent detection engineering in a continuous offense-defense testing loop. In an isolated environment modeled on NVIDIA accelerated computing infrastructure, Falcon sensors captured telemetry while Nemotron 3 Ultra orchestrated the defensive workflow and a customized Nemotron 3 Super generated or repaired detections. The harness grounded agents in sensor schemas and attack traces, then applied linting, replay, structured correction, and independent review before deployment. Backtesting raised mean detection of the recorded attack from 16.5% to 41.9% with the optimized open pipeline across independently seeded sessions. In live-fire tests, five of 11 open detections detected an unseen attack, three qualified as gold, and those three covered all eight attacks; the authors call the result a directional case study because it used one scenario family and limited benign traffic.
How Uken Games reduces observability costs by 87% with ClickHouse
Uken Games replaced Datadog with an open-source observability stack built around ClickHouse to monitor backend services supporting mobile games with millions of players. The migration addressed rising Datadog costs and vendor lock-in while targeting pragmatic feature parity across metrics, traces, and logs. OpenTelemetry collectors run in two layers: lightweight agents on ECS instances pre-filter traces, while an autoscaling gateway batches and buffers writes to SigNoz on ClickHouse; managed Prometheus handles metrics, CloudWatch retains logs, and Grafana provides dashboards and alerts. Uken samples ordinary traces, retains errors and high-latency requests, computes alert metrics without sampling, and limits retention to two weeks. The resulting system stores all traces on one ClickHouse node using about 170 GB, operates without migration downtime, and costs about $12,000 annually—an 87% reduction.
Top topics
- Architecture834
- Performance816
- LLMs660
- Machine Learning622
- Open Source563
- AI Agents553
- AI470
- Developer Experience448
Weekly digest
The best engineering reads, once a week.