Engineering & Perspectives
The Glintbase Blog
Technical essays, architectural deep-dives, and engineering perspectives on software design, context engineering, and autonomous AI agent operability.
ARS 3.3: Gates, Dual-Era MCP and What Changed on the Leaderboard
ARS 3.3 caps the score of sites agents can't read, stops scoring llms.txt for merely existing, probes MCP servers in the 2026-07-28 protocol and finds Server Cards where the final spec puts them. We scanned all 103 leaderboard companies with the release code: the average fell from 51.5 to 49.1, and 28 sites hit a gate.
What the llms.txt Evidence Actually Says
Three large studies found no link between serving llms.txt and AI citations, and most llms.txt files are never requested. Yet agents that find one through a link do use it. We read the evidence, checked 102 developer companies ourselves, and changed how the Agent Readiness Score treats the file.
Agent-Ready Means Governed, Not Just Reachable
Two large October 2026 surveys found that most enterprises already run AI agents but few can trace, see or stop them, and a third have found MCP servers nobody approved. Readiness has a second half: agents that can be identified, scoped and stopped. Here is what of that a site can show from the outside, and what it can't.
How ARS Compares With AgentReady, Ora and the IETF Draft
Three public definitions of agent readiness appeared in October 2026. We mapped every Agent Readiness Score check against AgentReady.org's requirements, Ora's four layers and the IETF individual draft, and explain where our weights come from, including what we borrowed and what we still need to prove.
The MCP Auth Fix That Ships Turned Off
Between 28 September and 15 October 2026 the official MCP SDKs published 13 security advisories. The fix for servers accepting tokens issued for another service is opt-in in both the Python and TypeScript SDKs: upgrading alone changes nothing. Here is what to upgrade to and the two changes that actually turn the check on.
Your robots.txt Lets AI Agents In. Your Edge Might Not
We fetched the same public pages on 150 popular websites as a declared AI agent, an unlabelled script and a headless browser. Saying we were an agent cost nothing, yet on 16.8% of sites whose robots.txt allowed us the edge still stopped us. Here is what we found, and what site owners should check.
What 3,000 MCP Servers Told Us About the Basics
We probed a random sample of 3,000 remote endpoints from the official MCP Registry using only protocol requests. Most work, auth is mostly done right, tool schemas are strong, but one in four servers can't tell an agent that a call failed, and only 13% speak the current protocol revision.
Why Agent-Readiness Scores Disagree
We ran three public sites through four agent-readiness scanners, our own included. The same site scored anywhere from 47 to 96 depending on the tool, and every tool ranked the three sites differently. Here is what each one measures, and how to read any readiness score.
Why We Rebuilt the Agent Readiness Score Around Evidence
ARS 3.2 replaces existence checks with verified evidence: live MCP probing, real OpenAPI inspection, honest 'unknown' states and one scoring engine across the website, CLI and MCP server.
We Audited Our Own Research. Here's What We Got Wrong
A line-by-line audit of Sandboxes Edition 1 and the State of Agent Readiness 2026 report found pass rates, human stops, traps and headline scores that the data did not support. Here is every major correction, and why we published errata.
What Auditing Our Own Stack Taught Us About Building Agent Tooling
An engineering postmortem from the Glintbase overhaul: fabricated fallback scores, permissive auth, rate-limit bypasses, scorecard takeovers and flaky tests across our scanner, site and CLI, and the principles we are keeping.
Best Cloud Sandboxes for AI Coding Agents (2026 Benchmark)
An empirical comparison of Modal, E2B, Daytona, boat.dev, and Vercel Sandbox for autonomous AI coding agents based on 30 live trials.
E2B vs. Modal vs. Daytona: Which Cloud Sandbox is Best for AI Agents?
A deep architectural comparison of E2B, Modal, and Daytona for autonomous AI coding agents: token efficiency, latency, microVM isolation, and SDK design.
How to Prevent Leaked MicroVMs & Compute Sprawl in AI Agent Workflows
How to prevent zombie cloud sandboxes and runaway infrastructure bills during autonomous multi-turn AI coding agent sessions.
Why Python Decorator Runtimes Score Low on Machine Discovery but Win at Runtime
The paradox between static documentation readiness and runtime execution: why decorator-driven cloud SDKs can score low on machine discovery yet perform well in live agent trials.
The Signed Preview URL Trap: Why Instant Ingress Tunnels Fail for AI Agents
Why instantaneous signed preview URLs in cloud sandboxes can cause transient HTTP 502 errors and hallucination loops for autonomous coding agents.
What is Integration Tax? The Hidden Infrastructure Cost of AI Agents
How autonomous AI coding agents burn extra tokens, tool turns and time navigating developer infrastructure, and how to measure Integration Tax.
The $433k Token Tax: Why Bad Docs Drain AI Infrastructure Budgets
How defective docs can cost $433k per year in modelled token overhead per 100,000 agent sessions a month. Learn to calculate and eliminate the token tax.
The Agent Resilience Paradox: Why 'Working' Software Fails AI Agents
Discover why software that works for humans regularly fails AI coding agents, and how the Agent Readiness Score captures what traditional QA misses.
Beyond /llms.txt: The Full Machine-Readable Stack for AI Agents
Explore the full machine-readable stack: llms.txt, mcp.json, and OpenAPI schemas that transform software into a first-class AI-operable surface.
State of Agent Readiness 2026: What We Found Across 100 AI and Developer Platforms
Findings from a simulated, deterministic agent traversal of 100 AI and developer infrastructure platforms across 10 benchmarks, showing the state of machine readiness in 2026.
How to Gate CI/CD on Agent Readiness: A Practical Glintbase CLI Tutorial
Use the Glintbase CLI to enforce Agent Readiness checks in CI/CD, blocking releases when machine-facing documentation falls below target thresholds.
Agent Readiness: The Missing Metric in Modern Software
Why software teams need to measure how well their products work for AI agents, not just humans.
Why AI Doesn’t "Use" Software the Way Humans Do
Why agents retrieve, traverse, execute, and infer differently — and why software must be designed for that reality.
The Internet Was Built for Humans. AI Agents Need a Different Interface.
Why software needs an agent-facing layer alongside human UIs, and how structured context and machine entrypoints unlock autonomous AI operability.
What is AI Readiness? The New Benchmark for Codebases
Understanding how machine-operable software is changing the meaning of good documentation, good architecture, and good developer experience.
Docs Drift Is Killing Your AI Coding Productivity
Why outdated documentation quietly breaks AI coding agents, and why static analysis plus runtime validation is becoming essential.
The Future of Agent-Operable Software
Why the next generation of software will need machine-readable documentation, context layers, and operational structure for AI agents.
Why Most RAG Systems Fail Developer Documentation
Why retrieval alone is not enough for technical docs, and why structural context beats simple vector search.