Most teams are buying AI developer platforms before they know how to evaluate them. This checklist cuts through the noise with 15 questions across five context categories — the answers reveal whether a platform treats context as infrastructure or is just a generic LLM with a polished UI. Bring it to your next vendor call.

Before deploying AI agents, it's worth thinking about how much of your company workflows, structure and process is actually documented in your software catalog and elsewhere. It's a direct measure of how legible your engineering org is to a machine.

AI agents fail in production engineering environments not because of model capability, but because they lack grounding data. This post breaks down the five specific context categories: service ownership, deployment state, incident history, tech standards, and runbooks. These determine whether an agent returns useful output or generic noise, and why most orgs already have this data but haven't made it queryable.

The PocketOS database deletion was three missing infrastructure layers, not model failure. Scoped tokens, API-level confirmation, and offsite backups: enforcement the model can't reason past.

The first wave of agent building left a governance gap. Here's why runtime governance - policy enforcement, context quality, and human oversight - is the fifth infrastructure bet.

Shared definitions for platform engineering teams working with AI agents: agent, harness, context layer, context lake, business context layer, minimum viable context, MCP, MCP Server, MCP Gateway, and tools.

Most enterprise AI deployments fail in production because the gap is context, not the model. The four motions of a context layer - ingest, organise, retrieve, refresh - explained for platform teams.

AI coding tools fail in production because they don’t have access to the context they need. This article explains the operational context layer your AI assistant is missing.

AI coding tools fail in production engineering orgs because they lack structured system-level context. Here's how to build the metadata infrastructure that produces reliable AI suggestions.

Context engineering is the architectural discipline that helps prevent enterprise LLM hallucinations. Here's what it covers, why most teams skip it, and how to audit your readiness before shipping.

August is a time for summer holidays and beaches and getting vaguely sunburnt but in a nice way. Or, if you're Roadie folks: building AI tools for Backstage. This month we integrated some MCP server tools, pushed out API token permissions for everyone, made mass updates possible, and moved OpenSearch to GA.

AI is already shaping internal developer portals, with platform teams testing RAG search, service ownership bots via Slack, and agent-driven automations. From the front lines, the same hurdles keep surfacing though: noisy context, hallucinations, fragmented tooling, and no governance. We take a look at how AI is affecting IDPs, and the challenges faced by the platform teams who run them.

AI cometh... In July we launched our MCP servers to help pull Roadie data into you LLM-powered IDEs and MCP Clients, took a step towards fully integrating RAG AI into Roadie Search, allowed customers to vibe code their own UIs and launched the Scheduler to improve transparency of the inner workings of Roadie. Oh, and updates to our Launch Darkly plugin.