AI agents: start here.
Agents are ordinary software with an unusual failure profile. These posts cover how they are built, where they go wrong in production, and how to contain the code they write and run.
I build PandaStack, a sandbox platform for agent code. The sandbox and isolation posts say so where it appears.
Kept current
AI Agent Security Incident Log
A dated, sourced log of publicly documented security incidents involving AI agents, coding agents, MCP servers and the AI software supply chain since 2025, split into breaches and CVEs, and kept current as new advisories land.
Open →Tracker · updated Nov 12, 2026AI Model Release Tracker
A living tracker of notable large language model releases since July 2026, hosted and open-weight: release date, vendor, context window and licence or access terms, each row linked to a primary source and kept current.
Open →Read in this order
- What Are AI Agents? Workflows, ReAct and Real Success Rates (2026)
AI agents explained for engineers in 2026: Anthropic's workflows vs agents split, the ReAct loop, tool calling, planning, memory and multi-agent designs, measured success rates from OSWorld (66.3%) to TheAgentCompany (30%), and when not to build one.
- AI Agent Frameworks Compared 2026: LangGraph vs CrewAI vs ADK
Ten AI agent frameworks compared: LangGraph, CrewAI, OpenAI Agents SDK, Claude Agent SDK, Google ADK and more, with versions, licences and how to pick one.
- Why AI Agents Fail in Production: 5 Real Incidents
What goes wrong when AI agents run in production: Replit, Kiro and PocketOS incidents, 14 failure modes from 1,642 traces, and the guardrails that stopped them.
- Multi-Agent Systems in 2026: +90% on Research, -70% on Planning, 15x Tokens
When multi-agent systems beat a single agent in 2026: Anthropic's 90.2% research gain at 15x the tokens, Google's 260-configuration study (+80.8% to -70%), Cognition's single-writer rule, the MAST failure taxonomy, and how to evaluate before you split.
- Context Engineering Explained (2026): Why 1M Tokens Still Isn't Enough
Context engineering in 2026, explained by an infrastructure engineer: why it replaced prompt engineering, what context rot research shows, and how retrieval, tool definitions, memory, compaction and sub-agents keep an agent's window small and useful.
- Agent Memory in 2026: Vendors Claim 92.5 on LoCoMo, a Plain Filesystem Scored 74, and Every Memory Product Lost to BM25
Agent memory systems in 2026, kept honest: Mem0's 66.9 on LoCoMo in its own paper versus 92.5 on its managed platform a year later, Zep's 75.1 rebuttal and the three bugs behind Mem0's 66.0 for Zep, Letta's 74.0 with only a filesystem, MemoryAgentBench putting every commercial system under 45 on retrieval against BM25 at 60.5, Hindsight's 89.6, first-party memory from ChatGPT, Claude and Gemini with dates, both labs' compaction APIs, what re-reading 115k tokens costs at cache-read prices, a Postgres and pgvector layer with supersession, and why memory poisoning is still unfixed.
- gVisor vs microVM vs Containers for AI Agent Isolation
Which isolation model to use for AI agent code execution: containers, gVisor and Firecracker microVMs compared on five axes by someone who runs a fleet.
- Who Is This Agent? Non-Human Identity in 2026, From the RFCs to the Breaches
Machines outnumber humans 109 to 1 in the enterprise directory, the OAuth working group has 70 drafts in flight, MCP just deprecated dynamic registration, and an agent with an over-scoped token deleted a production database in nine seconds. What actually standardised this year, what shipped, and the delegation chain I would build.
Guides and fixes
- How to Build an AI Agent With Tool Calling in Python (2026)
- How to Add Human-in-the-Loop Approval to a LangGraph Agent (2026)
- How to Build Agents With Claude Agent SDK Subagents (Python, 2026)
- How to Run an AI Coding Agent in GitHub Actions Safely (2026)
- How to Write AGENTS.md and CLAUDE.md Files Agents Follow (2026)
- How to Block Dangerous Commands With Claude Code Hooks (2026)
- How to Defend LLM Apps Against Prompt Injection: 8 Steps (2026)
Latest on AI agents
Anthropic Lets You Self-Host Claude Managed Agents, in Private Beta
On 12 November 2026 Anthropic opened a private beta of self-hosted Claude Managed Agents: the API server, agent loop, session history and workers run on your infrastructure, and only model calls leave. What changes for platform teams.
6 minNov 12, 2026MCP Changelog Tracker: Spec Versions, SDKs and Registry Status
A living tracker of the Model Context Protocol for engineers who run MCP servers: every specification revision, the latest release of each official SDK, official registry status and notable security advisories, each row sourced and kept current.
10 minNov 10, 2026Why RAG Fails in Production (2026): 7 Failure Points and Real Fixes
Why RAG fails in production in 2026: chunking that falls off a 2.5k-token context cliff, retrieval misses that contextual retrieval plus reranking cut by 67%, stale and post-filtered vector indexes, permission leaks like EchoLeak, long context and agentic retrieval trade-offs, and evaluation gaps.
11 minNov 10, 2026How to Run MCP Servers in Docker Securely (2026 Guide)
Run MCP servers in Docker without handing them your laptop: Docker MCP Toolkit profiles and the MCP Gateway, a hand-built stdio container with a read-only filesystem and no network, keychain secrets, Claude Desktop and Claude Code setup, and fixes for common errors.
12 minNov 10, 2026Plugin4Shell: Plugin SHA-Pin Bypass in Claude Code, Codex, Copilot, Gemini CLI
Air Security's Plugin4Shell showed that four AI coding agents trusted a pinned plugin SHA without checking it. Claude Code 2.1.179 and Codex 0.146.0 are fixed, Gemini CLI never will be. Versions, timeline and what to do.
6 minNov 10, 2026Operation Snowball: Jailbroken DeepSeek Agent Breached a Telecom and 13 Firms
Malbear Labs found an open directory holding 70+ transcripts of an autonomous DeepSeek-based agent that spent two weeks inside a European telecom and 13 hosted organisations, reaching millions of patient records.
6 minNov 10, 2026OpenAI GPT-6.1 Terra: Computer-Use Model at $1/$5, Half the Price of Sol
On 26 October 2026 OpenAI launched GPT-6.1 Terra, a mid-tier model for computer use and coding priced at $1/$5 per million tokens. Its benchmark claims, how it fits the GPT-6 lineup after the cancelled GPT-6.1 Astra, and what engineers should test.
6 minNov 10, 2026OpenAI DevDay 2026: Agents API Gets Computer Use, Codex Moves to the Cloud
OpenAI DevDay on 29 September 2026 added computer use and multi-agent features to the Agents API, cloud Codex, a Decisions API preview, a $500 Pro plan and Bedrock Managed Agents on AWS. What engineers should note.
6 min