Documentation
Prose references, guides, and conceptual content.
ARS 3.3 Standard • Applied AI Research Lab
Software Was Built for Humans.The Next Users Are AI Agents.
Glintbase is the research and engineering lab developing the Agent Readiness Standard (ARS 3.3), multi-agent flight simulation, and autonomous self-healing infrastructure for developer software and APIs.
The way software is consumed has fundamentally changed.
For decades, every assumption baked into software design was built around a human on the other end. That assumption no longer holds.
Human user
AI agent
01
Browse
Retrieve
02
Search
Reason
03
Click
Execute
04
Read
Infer
05
Scroll
Consume context
Most software was designed for none of this. The documentation, APIs, onboarding, and support surfaces that exist today were built entirely for human consumption — and they break in predictable ways when AI agents attempt to use them.
Current software breaks when AI agents try to use it.
These are not edge cases. They are the predictable, structural consequences of building software that was never designed for machine consumption.
01
AI wastes context.
When documentation is unstructured, narrative, or formatted for human reading, agents consume enormous token budgets attempting to extract actionable information. Most of what they consume is irrelevant to their task.
02
AI encounters dead ends.
Incomplete API references, broken cross-links, undocumented parameters, and missing authentication flows cause agents to stall mid-journey. Without recovery paths, they give up or hallucinate continuations.
03
AI hallucinates.
When structural information is absent or ambiguous, agents fill gaps with plausible-seeming fabrications. This is not a model failure — it is a documentation failure. Missing context forces inference.
04
AI repeats work.
Without persistent, structured knowledge representations, every agent session begins from scratch. Agents re-parse the same surfaces, re-resolve the same ambiguities, and re-discover the same information.
05
AI consumes excessive tokens.
Prose-heavy documentation, inconsistent structures, and redundant content dramatically increase the token cost of operating software. At scale, this translates directly to operational expense and latency.
An intelligence layer for software.
Glintbase sits between your existing software and the AI agents that need to operate it. It continuously extracts, structures, and optimises the knowledge those agents require.
Runtime analysis
Evaluates software surfaces as they actually behave, not as they are documented to behave.
Graph intelligence
Maps relationships between entities, surfaces, endpoints, and concepts across an entire codebase.
Semantic understanding
Derives meaning from structure — not just text — to build representations agents can navigate.
Journey simulation
Traces the paths an AI agent would follow to complete real tasks, identifying failures before they occur.
Intelligence pipeline
Every surface your software exposes to agents.
Scanner analyses ten distinct software surfaces. Hover any card to see what we look for.
APIs
REST, GraphQL, and RPC endpoint surfaces.
SDKs
Client libraries, type definitions, and wrappers.
GitHub
READMEs, wikis, and repository structure.
Support Centers
Help articles, FAQs, and troubleshooting guides.
Onboarding
Setup guides, quickstarts, and first-run experiences.
Authentication
OAuth, API keys, and identity flows.
MCP
Model Context Protocol servers and tool manifests.
OpenAPI
OpenAPI 3.x and Swagger specification files.
CLI
Command-line interfaces and developer tooling.
Soon
More coming
LLMs.txt, changelogs, runbooks, internal wikis.
Measure agent readiness. For free.
Run Scanner against any public URL. In under a minute, you receive a readiness report across four layers (Discovery, Access, Usability, Payments), with every check's evidence. No account required.
- ARS 3.3 composite score (137-check taxonomy)
- Multi-agent Flight Simulator (Claude Code, Cursor, Perplexity)
- Anti-SPA 404 canary leak verification
- Streamable HTTP Model Context Protocol (MCP) testing
- WorkOS auth.md machine authentication check
- Token Tax & context window dollar accounting
What a report showsARS 3.3 · 137 checks
- A 0–100 score and grade, stamped with its score version
- Layer scores, and how much of each layer could actually be verified
- Every check with its evidence: pass, warn, fail, or skipped when it couldn't be verified
- Blockers that cap the score: bot challenges and JavaScript-only homepages
- Ranked fixes, with the files and commands to change
Real scans. A published rubric.
What the scanner found on the 102 reachable companies of our public leaderboard, and how the 137-check ARS 3.3 score is computed. Every number here comes from a saved scan or the open-source check registry.
leaderboard_rescan · ars-3.3.1
Scanned 2026-10-25Companies scored
102
of 103 on the leaderboardAverage score
48.3
median 47 / 100Hit a gate
28
17 bot challenge or 403 · 11 JavaScript-onlyGrades
A 2B 24C 16D 43F 17
What the scanner verified
Homepage served to the scanner (no bot challenge or 403)83 of 100
Content readable without JavaScript49 of 83
Missing pages return a real 40481 of 100
Markdown returned when asked for33 of 102
OpenAPI spec found19 of 102
MCP server found57 of 102
MCP server speaks the current revision (2026-07-28)24 of 32
MCP errors reported in-channel32 of 32
“Of” counts the sites where the check could be verified; the rest were skipped (for example, MCP checks on sites without a reachable server). ARS 3.3 scan (score version ars-3.3.1) of the public leaderboard on 2026-10-25, saved scores only.
Score your own site:
Open the leaderboardnpx @glintbase/cli audit https://your-site.comFor teams who need certainty.
Deep Audit is an expert-led engagement that delivers a complete understanding of your software's agent readiness — and a precise roadmap to improve it.
Multi-Agent Flight Simulation
Empirical multi-turn tracing across Claude Code (200k), Cursor (128k), and Perplexity (zero-JS) personas.
E2B Runtime Sandboxed Execution
Executing documentation code blocks and API calls inside isolated Linux micro-VMs to verify parameter truth.
White-Box Offline AST Inspection
Direct AST code analysis of Next.js middleware, HTTP headers (Vary: Accept), and API handlers from source.
Counterfactual 'What-If' Proof Card
In-memory ArsSandbox re-scores each proposed fix and shows the measured change before anything is written to disk.
Anti-SPA 404 Canary Leak Prevention
Active detection and remediation of soft-200 HTML shells that induce severe agent hallucination loops.
Executive Governance & Compliance Report
Board-ready audit encompassing OWASP LLM Top 10, ISO/IEC 42001, and WorkOS auth.md specification parity.
Systems intelligence. Built for AI agents.
Empirical evaluations and benchmarks investigating how autonomous AI coding agents navigate, integrate, and execute across developer infrastructure.
Date
Category
Title
Oct 13, 2026
Agent Readiness
The Silent Wall: How Bot Defences Treat a Self-Declared AI Agent
Oct 13, 2026
Autonomous Infrastructure
Remote MCP Server Reliability: A Census of the Official MCP Registry
Oct 7, 2026
Autonomous Infrastructure
The Glintbase Integration Index: Cloud Sandboxes for Autonomous AI Coding Agents
Engineering insights. Built for builders.
Technical essays, architectural breakdowns, and engineering perspectives on software design, context engineering, and autonomous AI agents.
Built in the open.
The core of what Glintbase is building is open source. We believe agent readiness infrastructure should be inspectable, forkable, and community-owned.
@glintbase/cli
stable
Glintbase Agent Harness (ARS 3.3). Autonomous auditor with 137 checks, Multi-Agent Flight Simulator, AST code doctor, and CI drift shield.
TypeScript
glintbase/glintscanner
stable
Open-source scanner engine with E2B isolated runtime sandboxing, 3D knowledge graph, and deep crawl extraction.
TypeScript
@glintbase/mcp
stable
Model Context Protocol server with 17 production tools and 9 bundled skills for Claude Code, Cursor, Windsurf, and Antigravity.
TypeScript
glintbase/action
stable
GitHub Action PR Drift Shield. Automatically gates pull requests on ARS thresholds with automated PR Markdown scorecard comments.
TypeScript
Vision
Software increasingly serves machines.
Software is increasingly operating in a world where its primary consumers are not humans.
AI agents — systems that retrieve, reason, execute, and infer — are becoming active participants in every software workflow. They read documentation. They call APIs. They follow onboarding flows. They use CLIs, SDKs, and support surfaces.
This is happening now. And virtually no software was designed for it.
The gap between what software exposes and what AI agents need to operate it effectively is one of the most consequential infrastructure problems of the next decade.
Glintbase exists to close that gap.
We are building the intelligence layer that enables software to become understandable, navigable, and operable by the machines that increasingly depend on it.
This is not a documentation product. It is not an AI writing assistant. It is infrastructure — the kind that sits underneath, does its work quietly, and becomes foundational.
The companies that build this layer into their software now will have a compounding structural advantage as the agentic era develops.
We are building that layer.
Get Started
Help shape the agentic internet.
Run a free scan on any URL. Or join the waitlist to be part of what we build next.