Non-profit community publishing original, research-backed articles on building production AI agents. We cover the tools, frameworks, protocols, and security practices that define the agent engineering field in 2026.
64 published articles across 8 core topics, written by practitioners, not marketers.
What We Cover
AI Agents & Autonomous Systems: architecture, deployment, and real-world case studies
Model Context Protocol (MCP): the open standard connecting AI to tools and data
Agent Frameworks: LangChain, CrewAI, AutoGen, Google ADK, AWS Strands, and emerging tools
Prompt & Context Engineering: the skill shift from prompt crafting to context architecture
Agent Security & Guardrails: OWASP Top 10, sandboxing, trust boundaries
Multi-Agent Collaboration: orchestration, benchmarks, team coordination
AI Infrastructure & Tooling: Docker, CI/CD, observability for agent systems
Research & Benchmarks: academic papers, industry reports, competitive analysis
Latest Articles
The Key Was Deleted in 3 Hours. It Worked for 4 Months.: A LiteLLM admin key was deleted from a public repo in three hours but never revoked. 136 days later it led to a signed, malicious release. We mapped the fixes. [security] 2026-11-09
The Protocol Costs 7 Milliseconds. The Planner Costs 3.2x.: A multi-agent protocol added 7.2 ms per handoff. Adding a planner and reviewer cost 3.22x and solved no more tasks. We reran the data to find the saving. [multi-agent] 2026-11-06
OpenAI Monitors Internal Codex. Nobody Gets Paged.: OpenAI monitors all of its internal Codex traffic with paging turned off. METR's red team and a DNS incident show why detection was never the slow part. [security] 2026-11-02
OpenAI Now Runs Your Agent Harness. Read the Defaults.: The harness is the part of an agent nobody demos. It is the loop that calls the model, runs the tools, decides what to keep when the context fills, restarts the work that crashed at 3am, and holds a session together across hours. [announcement] 2026-09-21
The UN Shipped an MCP Server That Admits What It Cannot Do: Somewhere in the documentation for the United Nations' new MCP server there is a note that almost no vendor would publish. One of the six tools, getmultientityobservations, "has no applicable target on this instance." [mcp] 2026-09-20
The Agents Called It evil.rb. OpenAI Calls It Benign.: Someone published a Ruby package called zzsouthrunner in May 2026. At the top of its payload sat a comment that saves everyone a lot of forensic work: malicious crawler/exfil for Southwark Jan 2026 docs via rubydoc.info worker The package was one of more than 2,000 uploaded... [security] 2026-09-20
Compaction Summaries: The Rare Jailbreak, the Common Lie: An agent building a financial model could not find the historical data it had been asked to use. So before its context window ran out, it wrote a note to the next context: create a tab called Historical Data with "reasonable 2024 historical data," and "be transparent only if... [research] 2026-09-18
Two Conferences, One City, One Day: On Tuesday 29 September, two AI events open in San Francisco on the same day, at opposite ends of the city's waterfront. At Fort Mason, OpenAI holds DevDay: a single day, $650 to attend, with the opening keynote livestreamed. [event] 2026-09-11
MCP Shipped Identity for Humans. Agents Are Still a PR.: The MCP roadmap published on 22 August makes a clear promise. Under the heading "Agent identity and enterprise-ready security," the lead maintainers write that they want MCP servers "to have a standardized way to recognize and trust those agent identities, built on existing... [mcp] 2026-09-11
18 of 30 Agents Picked the Same Branch Name: Thirty agents were turned loose on a shared software project. Eighteen of them created a git branch with the exact same name: mvp-game-loop. Nobody coordinated that. Nobody told them what to call it. [multi-agent] 2026-09-09
Microsoft's Own README Says MXC Is Not a Security Boundary: Buried in the README of Microsoft's new agent sandboxing project is a sentence that no one writing about it seems to have quoted: "no MXC profiles should be treated as security boundaries currently". [infrastructure] 2026-09-08
MCP-Atlas Says 44.5% and 82.2%. Ask Which Number.: MCP-Atlas is becoming the benchmark people quote when they want to say a model is good at tool use. Meta cited it in the Muse Glimmer release, reporting 75.5%. It is a good benchmark, built with more care than most. [mcp] 2026-09-02
Agreement Is Not Accuracy: Four Papers on Agent Debate: Multi-agent debate is sold as an accuracy technique. Put several models in a room, let them argue, take the consensus, get a better answer than any one of them would have produced alone. [research] 2026-08-31
HTTP 402 Sat Unused for 29 Years. Agents Changed That.: When HTTP/1.1 was specified in 1997, the authors reserved status code 402 for "Payment Required" and then left it alone. For nearly three decades it has been the web's most famous piece of unused plumbing: a slot held open for a payments layer the web never got. [news] 2026-08-26
Cloudflare Bets Your Agent Can Pick Its Own Backend: Most agent sandbox products are sold on developer experience: here is an isolated place to run untrusted code, here is how you attach it. Cloudflare's @cloudflare/computer, released as an open source preview on 3 August by Matt Carey and Aron Carroll, is sold on something... [infrastructure] 2026-08-24
LiteLLM and SANDCLOCK: The Scanner Was the Attack Path: Every stolen-secret file in the LiteLLM breach archive is named the same way: a timestamp, then 127.0.0.1. That address is the most informative detail in the entire incident. It means the secrets were not intercepted in transit. [security] 2026-08-23
185,616 Stars, 811 Watchers: Read the Second Number: DeepSeek Harness went public on 13 August at 11:56 UTC. Nine days later the repository shows 185,616 stars and 20,555 forks. That is one of the steepest adoption curves GitHub has recorded for a developer tool, and it is the number every write-up has led with. [tool] 2026-08-23
Meta's 30B Agent Model Fits in 24GB. Read the Attention.: Meta released Muse Glimmer on 10 August: roughly 29.6 billion parameters including the vision encoder, Apache 2.0, and quantized small enough to run inference on a machine with 24 to 32GB of memory. Half a million downloads in the first month. [announcement] 2026-08-22
Gemini 3.7 Flash Is Not a Price Cut. It Is a Countdown.: The most useful number in Google's Gemini 3.7 Flash launch is not $0.75. It is 1 January 2027. The model shipped on 13 August, described by Google as "our most intelligent workhorse model yet for coding and agents". The coverage that followed led on price. [news] 2026-08-21
Machine-Readable Access
Our content is fully open to AI agents and developers: