Thu 10 Sep · 17:00 CEST · onlineCommunity Town Hall: the roadmap, the three colours, and a live provisioning demoAgenda and registration →

Engineering Blog

Empirical benchmarks, architecture deep-dives, and research on autonomous infrastructure engineering with the Colors SDK.

Six Machines and a Contract Nobody Told the Gates About

Six Machines and a Contract Nobody Told the Gates About

We split Langfuse v4 across six Vultr machines — Neon for Postgres, a Redis host, three ClickHouse replicas, the app behind Cloudflare — and made every separation claim a gate. Eleven converges: a storage tier that left a table in the application database and broke Prisma, a ClickHouse backup that wrote 128 unnamed objects, and a v4 write mode that rejected our traces while answering 207. Then a restore that boots the pinned image and reads through the API.

#Field Report#Langfuse#ClickHouse#Neon#Acceptance#Vultr
Colors Community Town Hall, 10 September

Colors Community Town Hall, 10 September

A one-hour virtual town hall for the people building and running Package Skills: where the SDK is after the rebrand, how the three colours fit together, an agent provisioning an analytics and application stack live, the proposals on the roadmap, and open Q&A with the maintainers. Free, registration on Luma.

#Community#Town Hall#Roadmap#Package Skills#Live Demo
The Converge Passed in 614 Milliseconds

The Converge Passed in 614 Milliseconds

We put n8n's database on object storage — a colocated self-hosted Neon instead of a colocated Postgres — and it handled 7950 executions in five minutes at a p99 of 80 ms. Getting there took sixteen converges, and the most useful failures were not in the infrastructure. Four of our own acceptance gates were wrong, one reported a green deployment with nothing checked at all, and one accused a correctly hardened host.

#Field Report#n8n#Neon#Acceptance#Object Storage#Vultr
The Cluster Was Fine

The Cluster Was Fine

A three-node AutoMQ cluster whose disks are object storage took twenty-three converge runs. In most of the failures the cluster was healthy: the firewall passed ping while dropping every TCP connection, a marker claimed work that never happened, and three of our own acceptance gates were wrong — one of them reporting data loss that never occurred.

#Field Report#AutoMQ#Kafka#Acceptance#Object Storage#Vultr
The Docs Stop Where Operations Begin

The Docs Stop Where Operations Begin

Neon's engine is open source and genuinely good; its self-hosting story is a development fixture, and since July 2025 there is no versioned release to pin. We deployed it for real, measured the actual single-node durability bound, and left a Package Skill and a Context Skill behind as the starting point upstream doesn't ship.

#Field Report#Neon#PostgreSQL#Cloudflare R2#Package Skills#Context Skills
The Skill Your Build Leaves Behind

The Skill Your Build Leaves Behind

Introducing Context Skills: knowledge distilled from a verified build, routed by the error on your screen, honest about what verified it. Three are live in the Skills Catalog, with one workflow to create them and another to submit yours.

#Announcement#Context Skills#Agent Skills#Agentic DevOps#Knowledge#Catalog
How a Package Skill Gets Made

How a Package Skill Gets Made

Three clipped YouTube transcripts became a live NetBird Agent Network deployment in a day: a keyless LLM endpoint and an agent container that cannot reach the internet, proven by acceptance gates. The six-step workflow behind it, with the failures left in, and the closing move that makes the next run cheaper.

#Workflow#Agentic DevOps#NetBird#Claude Code#Adversarial Review#Package Skills
The Agentic DevOps Manifesto

The Agentic DevOps Manifesto

Thirteen theses on letting agents run infrastructure, written for the people who think that is a bad idea. Every claim is implemented in the getcolors repositories, and three of them exist to say what the approach costs rather than what it buys.

#Manifesto#Agentic DevOps#Desired State#Golden Tests#Package Skills#Guardrails
Park the Stochasticity

Park the Stochasticity

Anders Hejlsberg explained why the TypeScript team didn't let AI translate their half-million-line compiler to Go. Have it write a deterministic translator instead. Colors is that argument applied to infrastructure, where the agent writes the program once and the program runs production forever.

#Architecture#Agentic DevOps#Determinism#TypeScript#Package Skills#Golden Tests
Agentic Once, Deterministic Forever

Agentic Once, Deterministic Forever

An agent renamed a production analytics domain, fixed the bug the rename exposed, and moved the deployment to another cloud in ten minutes of downtime. The safety KPI is a ratio: deterministic tasks over agentic ones, and every session should move it up.

#Architecture#Agentic DevOps#Claude Fable 5#Rybbit#Vultr#Package Skills
Let Your Remote Coding Agent See Your Clipboard

Let Your Remote Coding Agent See Your Clipboard

A coding agent on a remote machine cannot see your screen. pngpaste, socat, a launchd agent, one ssh -R forward and a small skill give it your Mac's clipboard as a Unix socket. The recipe, and the two failures that each reported success.

#Architecture#Claude Code#SSH#launchd#Nix#Skills
You Probably Don't Need the Modern Data Stack

You Probably Don't Need the Modern Data Stack

For simple use cases, one $40 virtual machine running dlt, dbt, ClickHouse, PocketBase, and systemd can replace a four-figure SaaS data stack. A live single-host GitHub warehouse is the proof.

#Architecture#ClickHouse#dlt#dbt#PocketBase#Single Host
Both Agents Reported Success. Only One Checked.

Both Agents Reported Success. Only One Checked.

Two agents deployed PostHog to DigitalOcean. One had a curated skill and used 4.4× fewer tool calls. The more useful finding: the baseline shipped a deployment it never proved could store a single event.

#Benchmark#PostHog#Skills#ClickHouse#DigitalOcean#Agentic DevOps
Three Analytics Stacks, Three Agents, One Uncomfortable Finding

Three Analytics Stacks, Three Agents, One Uncomfortable Finding

Three isolated agents each built a self-hosted analytics platform as a Package Skill and deployed it live. All three reported success — and all three passed acceptance checks that could not detect a deployment serving 502s, backups that had never succeeded, or default credentials answering on the public internet.

#Benchmark#Umami#Rybbit#PostHog#ClickHouse#Verification
Benchmarking Claude Opus 5 on Autonomous Distributed Infrastructure: Building PostgreSQL & MySQL HA Clusters from Scratch

Benchmarking Claude Opus 5 on Autonomous Distributed Infrastructure: Building PostgreSQL & MySQL HA Clusters from Scratch

An empirical creation-effort benchmark evaluating Claude Opus 5 building and deploying 3-node PostgreSQL and MySQL High Availability clusters to DigitalOcean, with verified failover and continuous backups to Cloudflare R2.

#Benchmark#Claude Opus 5#PostgreSQL HA#MySQL HA#Patroni#Group Replication
Benchmarking Gemini 3.7 Flash on Autonomous Distributed Infrastructure: Building PostgreSQL & MySQL HA Clusters from Scratch

Benchmarking Gemini 3.7 Flash on Autonomous Distributed Infrastructure: Building PostgreSQL & MySQL HA Clusters from Scratch

An empirical creation-effort benchmark evaluating Gemini 3.7 Flash building and deploying 3-node PostgreSQL and MySQL High Availability clusters to DigitalOcean with continuous backups to Cloudflare R2.

#Benchmark#Gemini 3.7#PostgreSQL HA#MySQL HA#DigitalOcean#Cloudflare R2
Agentic DevOps Has a Compounding Advantage

Agentic DevOps Has a Compounding Advantage

Why agentic DevOps accelerates faster over time: every verified Package Skill creates an executable corpus of operational knowledge that reduces the reasoning and implementation required for future deployments.

#Agentic DevOps#Architecture#Colors SDK#Autonomous Systems#Compounding Advantage