Sources

27 ingested

Sources

Every source distilled into a cited learning note.

12-Factor Agents: Patterns of reliable LLM applications

Start here for agent fundamentals. An agent is a prompt + switch statement + context builder + loop - own all four. Tools are just structured JSON; what ships in production is micro agents (3-10 steps

Dex Horthy (HumanLayer) - AI Engineer (World's Fair 2025)
video

Building Closed-Loop Evals for a Multimodal Agent at Scale

Blueprint for evaluating an agent pipeline in production: log-first, per-stage metrics (routers as classifiers, generation pass@k, pairwise comparison), Swiss-cheese QA gates, and a closed self-tuning

Soumya Gupta & Jai Chopra (Uber, Computer Vision team) - AI Engineer (World's Fair)
video

Harness Design for Long-Running Application Development

A harness built, costed, then half-deleted on a newer model. Planner/generator/evaluator; the split exists to beat self-evaluation bias , not to add capability. Subjective quality becomes gradable by

Prithvi Rajasekaran (Anthropic Labs)
blog

OAuth 2.0 and OpenID Connect (in plain English)

The clearest explanation of the delegated-authorization substrate. OAuth exists to kill password sharing; its shape follows entirely from "the browser can talk to a human but cannot hold a secret". Fo

Nate Barbettini (Okta) / OktaDev
video

Don't Ship Skills Without Evals

The brain's first source on skills, and it arrives with third-party numbers. A skill is a three-layer cost ladder, not a document; its description is the trigger and causes 50%+ of all failures; lengt

Philipp Schmid (Google DeepMind) - AI Engineer World's Fair, Track 5 (Evals)
video

Dreaming: Better memory for a more helpful ChatGPT

The brain's first source on memory. Memory written during a conversation keeps that conversation's tense, so it decays into confident wrongness , not irrelevance. The fix is architectural: move the wr

OpenAI (no individual byline - the article credits "OpenAI")
blog

Memory and dreaming for self learning agents

The agent-platform half of the memory pair, and the independent counterpart to the OpenAI post above. Same architecture, same name, different vendor: agents write memory during work, and a decoupled b

Ravi (API knowledge team, Anthropic) - "Code w/ Claude" conference, Claude channel
video

LLM Wiki - "A pattern for building personal knowledge bases using LLMs"

The brain's first source on RAG, and it argues against it. Retrieval is stateless across queries - a synthesis question re-pieces the same fragments every time, at the moment the user waits, and keeps

Andrej Karpathy (personal gist, no affiliation stated, nothing being sold)
blog (public GitHub gist - a prose idea document, no code)

Inside the Microsoft Agent Framework: How we designed a layered SDK

A vocabulary, not a finding - and the brain's first flat cross-source contradiction. Three separable ideas: the agent loop (execution), workflows (orchestration), the harness (runtime capabilities) -

Shawn Henry, Principal Group Product Manager (the `commandline` engineering blog)
blog

Tool search: Finding the right tool at the right time

The brain's first real MCP source, and it is an article about something else. A tool catalog stops being schema management and becomes a search problem past ~10-15 tools. Replace tools/list with two m

Lisa Brown Jaloza - Command Line (Microsoft)
blog

How we built LangChain's agent-first data stack

The brain's first source on an agent reading a company's own proprietary knowledge - and the figure gives away more than the prose. The architecture diagram captioned as the "big architectural shift"

Emily Hawkins (data team, LangChain)
blog

Multi-tenant agentic AI system

The brain's first source on deploying agents across an organisation rather than building one - and the lesson is in its second half, not its diagram. Put the tenant boundary at the platform's own coar

Shivank Awasthi (Field Solutions Architect) + Utkarsh Bhardwaj (Technical Solutions Consultant, Agentic AI), Google Cloud Architecture Center - **21 further contributors listed**
blog (vendor reference architecture)

autoresearch - AI agents running research on single-GPU nanochat training automatically

The brain's first source on an agent that runs experiments unattended, and the first where the whole design is a permissions diagram. An agent gets one editable training script, a fixed 5-minute wall-

Andrej Karpathy (`karpathy`)
code

AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases

The brain's first paper source, its first measured attack, and its first source with nothing to sell. An agent's memory or knowledge base is an input nobody treats as one: retrieved records enter cont

Zhaorun Chen (U Chicago), Zhen Xiang (UIUC), Chaowei Xiao (U Wisconsin-Madison), Dawn Song (UC Berkeley), Bo Li (U Chicago + UIUC)
paper

Defeating Prompt Injections by Design

The brain's first gated defence , and the first source here whose security argument does not depend on the model behaving. The thesis is a reversal: stop making the model safe, and build a system in w

Edoardo Debenedetti (Google + ETH Zurich), Ilia Shumailov (Google DeepMind), Tianqi Fan (Google), Jamie Hayes (DeepMind), Nicholas Carlini (DeepMind), Daniel Fabian (Google), Christoph Kern (Google), Chongyang Shi (DeepMind), Andreas Terzis (DeepMind), Florian Tramèr (ETH Zurich)
paper

Stanford CS329A Self-Improving AI Agents - Part 1 - Course Overview

The brain's first academic source, and its first on a model improving itself. Self-improvement turns out to be loop closure rather than a technique : sample the model many times, keep what survives a

Azalia Mirhoseini (Assistant Professor, Stanford CS) + Akanksha Chowdhery (Adjunct Professor, Stanford; research at Reflection AI). Published on the Stanford Online channel
video

Stanford CS329A Self-Improving AI Agents - Part 2 - Test-Time Compute Scaling

The mechanism lecture 1 kept deferring, and the best thing in it is the demolition rather than the technique. Repeated sampling is lawful - coverage against sample count fits c = exp(a·k^b) from 70M t

Azalia Mirhoseini (Assistant Professor, Stanford CS), with Akanksha Chowdhery co-teaching. Published on the Stanford Online channel
video

Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection

The paper that named indirect prompt injection, and the source that moved agent-security to established . Its one load-bearing sentence: processing untrusted retrieved data is analogous to executing a

Kai Greshake (Saarland University + sequire technology GmbH), Sahar Abdelnabi (CISPA), Shailesh Mishra (Saarland), Christoph Endres (sequire), Thorsten Holz (CISPA), Mario Fritz (CISPA). **Greshake and Abdelnabi contributed equally**
paper

AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents

The field's reference benchmark for agent security, and the first artifact here built to settle arguments rather than win one. A user task and an attacker task run in the same stateful tool-calling en

Edoardo Debenedetti (ETH Zurich), Jie Zhang (ETH), Mislav Balunović (ETH + Invariant Labs), Luca Beurer-Kellner (ETH + Invariant Labs), Marc Fischer (ETH + Invariant Labs), Florian Tramèr (ETH)
paper

Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents

The best-evidenced source in this brain, and the second primary autonomous-research-loops was waiting for. Schmidhuber's Gödel machine rewrites itself only on a proof the change helps, which is why no

Jenny Zhang (UBC + Vector Institute), Shengran Hu (UBC + Vector + Sakana AI), Cong Lu (UBC + Vector + Sakana), Robert Lange (Sakana AI), Jeff Clune (UBC + Vector + Sakana + Canada CIFAR AI Chair). **Zhang and Hu are co-first authors; Lange and Clune co-senior**
paper

From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents

The first source anywhere here that enumerates how a poisoned memory write actually happens - and the third independent leg of claim 145, after S16 and S17. The economics first: prompt injection needs

Pritam Dash (Huawei Canada), Tongyu Ge (Huawei Canada), Aditi Jain (Huawei Canada), Tanmay Shah (University of Waterloo; work done during an internship at Huawei), Zhiwei Shang (Huawei Canada)
paper

Defending Against Indirect Prompt Injection Attacks With Spotlighting

The cheap defence most teams actually ship, and the analogy that names its own ceiling. First it measures the thing everyone tries: adding "do not obey instructions in the document" to the system prom

Keegan Hines, Gary Lopez, Matthew Hall, Federico Zarfati, Yonatan Zunger, Emre Kıcıman - **all Microsoft**
paper

Scaling AI Agent Infrastructure with the MCP Stateless updates

The brain's first primary MCP source, and the first anywhere to pin a spec version (2026-07-28, superseding 2025-11-25). The initialize handshake and Mcp-Session-Id are deleted and what they negotiate

Kurtis Van Gent (Senior Staff Software Engineer, Google Cloud Data) + Alan Blount (Technical Product Manager, Google Cloud AI) - Google Developers Blog
blog

Hermes Agent Architecture - Part 1: Gateway, Sessions, and the Agent Loop

The brain's first source on the operational runtime of an agent rather than the shape of its loop - and the first practitioner source here analysing somebody else's open-source project instead of sell

Vinoth Govindarajan ("The Agent Stack", Substack)
blog

Patterns for Building Cybersecurity Evals

The brain's first survey source, and the first to measure agents as attackers rather than as targets. Seven benchmarks read at once - Cybench, CVE-Bench, CyberGym, ExploitGym, ExploitBench, MHBench an

Eugene Yan (Ziyou Yan), eugeneyan.com
blog

LLM Knowledge Bases: a practical guide

The first independent instantiation of S8 (LLM Wiki), and the cleanest lesson this brain has on why that is worth less than it sounds. Ben Holmes read Karpathy's llm-wiki gist, built it, and demonstra

Ben Holmes (developer relations lead, **Warp**) - channel: AI Engineer
video (conference talk, AI Engineer World's Fair 2026, Track 3 "Memory & Continual Learning")

Scaling GitHub for your Agents

The brain's first operator-scale report on a public MCP server - roughly 7.34M tool calls a week, 101 tools arrived by open contribution, and agents got measurably worse at using GitHub. The finding u

Sam Morrow (leads development of GitHub's MCP server) - AI Engineer, recorded at AI Engineer Europe, London
video