opencode accepts --dangerously-skip-permissions as a hidden alias for --auto. What it approves, what deny rules still block, and the permission config in opencode.json.
- opencode
- coding-agents
- agent-security
WRITING
Notes from running agents
What actually ships, what gets cut, and why. Verdicts from daily use, not press releases.
Codex has no --dangerously-skip-permissions. Use --dangerously-bypass-approvals-and-sandbox, or -a never with a sandbox. Plus approval_policy, sandbox_mode and --approve-for-me.
- codex
- coding-agents
- agent-security
GA4 consent mode explained: basic vs advanced, what each sends before the click, and a store where Search Console counted clicks on days GA4 counted nothing.
- measurement
- ad-platforms
- field-notes
Choose RAG or fine-tuning by the failure you need to fix: missing facts, inconsistent behavior, or both. Includes a decision worksheet and a local retrieval fixture.
- rag
- fine-tuning
- llm-evaluation
Use n8n error workflows, node error outputs and bounded retries. A local receipt model shows why a timeout can repeat an action that already succeeded.
- automation
- agent-tools
- reliability
Configure the Claude Code sandbox, probe file and network restrictions, and check retries, excluded commands and tools that run outside its boundary.
- claude-code
- agent-security
- agent-ops
Understand Claude Code Channels, sender allowlists, session lifetime and permission relay. Use a local test plan before connecting an external chat platform.
- claude-code-channels
- claude-code
- agent-security
Calculate AI agent cost with tokens, tools, retries, failed tasks and operator time. A synthetic ledger shows what cost per successful task includes.
- ai-agent-cost
- agent-operations
- evaluation
Choose code, conversation or both in Claude Code rewind, check shell and subagent limits, and keep Git and external recovery separate from checkpoints.
- claude-code
- agent-ops
- git
Use Claude Code worktrees for parallel tasks, choose the right base, handle dependencies, review each branch, merge changes and clean up safely.
- claude-code
- agent-ops
- git
Choose a Claude Code plugin scope, inspect its components and context cost, diagnose loading, then update, disable or remove the correct installation.
- claude-code
- skills
- agent-ops
Build a small Codex skill with SKILL.md, choose its local scope, check discovery, and test explicit invocation without assuming tool access.
- codex-skills
- codex
- agent-workflows
GPT-6.1 Sol is here. Check dots access, ChatGPT Pro prices, Codex usage limits and the difference between faster modes and billing multipliers.
- codex
- agent-tools
- agent-ops
Claude Sonnet 5.5 keeps Sonnet 5 token prices. Check speed claims, API changes, and progress updates your agent interface could stop showing after migration.
- claude
- agent-ops
- coding-agents
Harrison Chase's own-your-intelligence framework, checked question by question against my own AI agent stack, with honest yes, no, and partial answers.
- ai-strategy
- agent-harness
- ai-agents
How to use AI agents for code review without losing control: specialized reviewer agents, a CI loop, fail-closed auto-merge gates, and verifying real behavior.
- ai-code-review
- agentic-ci
- code-review-automation
How prompt caching actually works under the hood, what is worth caching, when it does nothing for your bill, and the one-line check I run before turning it on.
- llm-cost
- prompt-caching
- claude-api
AI technical debt breaks into four types: data, model, prompt, and organizational. Here's what each looks like in practice, plus a self-audit checklist.
- ai-governance
- technical-debt
- ml-ops
Opus 5.5, GPT-6 Sol and Luna launched September 22, 2026. Compare API IDs, token prices, Work and Codex availability, and migration checks for agent builders.
- ai-models
- agent-tools
- claude
- codex
Context rot explained: why a large context window does not stop an agent from getting worse mid-session, plus how to spot the dumb zone and four ways out.
- context-window
- agent-context-management
- claude-code
- context-engineering
AGENTS.md and CLAUDE.md files bloat when every fix gets appended to them. A router-first structure, plus what OpenWiki automates, keeps the entry file small.
- agent-documentation
- context-engineering
- claude-code
Google Merchant Center feed audit: what a 548,575-item feed had wrong (empty MPN, constant availability, missing images) and which fixes actually moved.
- ad-platforms
- measurement
- field-notes
claude --resume: what a resumed session restores, what it drops, and what my cheat sheet got wrong
RUNNINGclaude --resume from the Claude Code docs: when to resume, when to /clear, what a resumed session restores and what it drops, from 215 transcripts on disk.
- claude-code
- agent-ops
- context-engineering
--dangerously-skip-permissions: what it skips, what still stops Claude, and the dead allowlist
TESTED--dangerously-skip-permissions: what it skips, what deny rules and hooks still block, and the matching switch in Codex, opencode, Cursor, Copilot and Gemini CLI.
- claude-code
- agent-security
- agent-ops
claude -p runs Claude Code with no terminal: prompt in, result out, JSON with cost and session ID. The flags my automations use and what --bare changes.
- claude-code
- agent-ops
- coding-agents
What is an agent harness: every piece of code around the model, the loop, the tools, the approvals and the log. Examples, harness vs agent, and the one I run.
- glossary
- agent-harness
- agent-ops
Claude Code context window explained from the docs: what loads at startup, what /context shows, when auto-compaction runs, what survives, and the rules I keep.
- claude-code
- context-engineering
- agent-ops
What is function calling: the model picks a tool and writes the arguments, your code runs it and sends the result back. The loop, the cost, and MCP.
- glossary
- agent-tools
- mcp
Claude Code memory has two parts: CLAUDE.md you write and auto memory Claude writes. Where each lives, the 200-line and 25 KB limits, and my 538 memory files.
- claude-code
- context-engineering
- agent-ops
What is a system prompt: the instructions an application loads before your message, what they can and cannot enforce, and how Claude Code builds its own.
- glossary
- prompt-engineering
- llm
Claude Code subagents run a task in their own context window and return a summary. What the docs guarantee, the two agent files I run, and one leaked .env file.
- claude-code
- agent-ops
- context-engineering
What AI guardrails are in operator terms: input filters, output checks and action gates, why the third one is the one agents need, and the three layers I run.
- glossary
- guardrails
- agent-security
Field notes on claude --chrome, the flag that hands Claude Code your logged-in browser: 27 screenshot timeouts, clicks that miss, and the gates I use.
- claude-code
- browser-automation
- agent-ops
ChatGPT Ads: a US test in February, self-serve in 52 countries by September. What Ads Manager Beta gives you: bids, context hints, Pixel and Conversions API.
- ad-platforms
- chatgpt
- measurement
What is RAG in AI, in plain terms: fetch relevant text from a source you chose, put it in the prompt, then generate. How it works, RAG vs MCP, where it fails.
- glossary
- llm
- rag
Before a tool call runs, a PreToolUse hook in Claude Code can deny it or ask me. My 79-line production guardrail, the ssh case rules miss, and its limits.
- claude-code
- agent-ops
- guardrails
- agent-security
A plain definition of what an AI agent is: a language model that picks tools inside a loop, with a stop condition and a human on the gates that matter.
- glossary
- ai-agents
The statusLine setting runs a script on every update. My claude code statusline reads only stdin: median 41 ms, replacing a tool that took 13-20 s a render.
- claude-code
- agent-ops
- context-engineering
A practical AI agent architecture diagram with six layers for goals, orchestration, context, models, tools, policy, and production observability in practice.
- ai-agents
- architecture
- agent-ops
A practical LLM evaluation method using representative test cases, deterministic checks, human review, calibrated model judges, and release gates that work.
- llm
- evaluation
- testing
Seven practical AI agent examples with their inputs, tools, stop conditions, human approval gates, and the failures each design must handle in production.
- ai-agents
- use-cases
- field-notes
A plain explanation of AI prompts, their main parts, prompt roles, practical examples, common mistakes, and how to test whether a prompt works in practice.
- glossary
- prompt-engineering
- llm
A practical MCP server security checklist covering trust boundaries, credentials, token handling, approvals, sessions, sandboxing, and logs for AI agents.
- mcp
- agent-security
- security
What is a context window in AI? Learn what fits inside it, how it differs from memory, why long context still fails, and how agents manage it in practice.
- glossary
- llm
- context-engineering
AI agent deployment needs more than a working prompt. Use this practical path through sandbox, staging, canary release, monitoring, rollback, and approval.
- ai-agents
- deployment
- production
What is an AI hallucination: a fluent, confident output that is not true. Why next-token prediction makes them inevitable, and which defenses work.
- glossary
Agents write code many times faster. Robert C. Martin says spend the saved time on testing rigor. Field notes on what that looks like when you actually do it.
- coding-agents
- agent-ops
What is a token in AI: the chunk a model reads and writes text in. Roughly three-quarters of a word, and the unit behind every context limit and every invoice.
- glossary
AI agent monitoring is not server monitoring. Field notes from running a 36-probe cockpit over an agent fleet: traces, review queues, judges, decision stamps.
- agent-ops
- production
- monitoring
Slack Code gives coding agents their own channels with diffs, previews and a self-archiving log. What it changes for permissions, prompts and who can build.
- coding-agents
- agent-ops
- agent-security
What is an LLM: a model trained on text to predict the next token, and the engine under chatbots and agents. How it works in practice and what people get wrong.
- glossary
One frontmatter line decides if Claude can run a skill on its own. disable-model-invocation on 55 real skills, 0 flagged then 8, plus Copilot, Cursor, Codex.
- context-engineering
- agent-ops
What is AI, in plain language: software that handles judgment calls, learned from data instead of programmed. What the term covers, and the myths around it.
- glossary
Anthropic cut over 80% of Claude Code's system prompt. The new CLAUDE.md best practices, tested against a real 482-line config and its 26 loudest rules.
- context-engineering
- agent-ops
A production-readiness checklist for AI agents: 12 gates across access, actions, failure and operations, each with a failure mode and a pass or stop test.
- agent-ops
- production
- guardrails
DeepSeek Harness ships under MIT with every capability as a plugin. What the docs let you swap, what they will not, and the tools left switched off.
- agent-harness
- coding-agents
- open-source
When an MCP server integration is worth writing yourself, what it takes, and why tool descriptions cost more thought than code. Notes from two servers I run.
- mcp
- agent-tools
- agent-security
Hosted and public MCP servers move the credential, the failure mode and the trust boundary off your machine. What I run, what broke today, what I check first.
- mcp
- agent-security
- agent-tools
An MCP server for your database gives an agent live SQL in minutes. Field notes on schema cost, read-only roles, and why my write path never goes through MCP.
- mcp
- databases
- agent-security
An MCP server for AI agents is judged differently than one for chat: tool count, description quality, error text, blast radius, and the check I run first.
- mcp
- agent-tools
- agent-security
An MCP servers list is only useful if it is short. The eight I keep connected, what earns a place, what gets one removed, and why directories do not help.
- mcp
- agent-tools
- context-engineering
What is an MCP server in AI: host, client and server explained, both transports, and a plain answer to where your data ends up when an agent calls a tool.
- mcp
- agent-tools
- context-engineering
- glossary
Agent harness design notes on Meta's Muse Code launch: an append-only event log, background agents that outlive the task, and an approval gate as a skill.
- agent-harness
- coding-agents
- context-engineering
Field notes, not a directory
RUNNINGWhy this blog only covers agent tools that carry real production work, and how the verdicts get decided.
- agents
- workflow