My agent failed the same caching refactor three times before I fixed how I was briefing it. Planning mode, browser access through Playwright MCP, and claude.md files are what turned that into one-shot work.
My agents decided their own heartbeat config was redundant and deleted it, twice. The tiered OpenClaw setup that came out of it: orchestrators for judgment, throwaway personas for tasks, and five Markdown files that hold an agent’s identity.
An agent that is 85% accurate per step is about 20% accurate over ten steps. Lusser’s Law explains why, the Replit database deletion shows what it costs, and a gatekeeper pattern in PHP puts human review in front of the irreversible actions.
WordPress.com added MCP support, so an AI agent can read and write the site directly. The same update brings a site-wide newsletter default, a command palette shortcut, button state controls, UPI payments for India, and a fixed subscriber chart.
Autonomous agents fail in ways that are expensive to clean up. This is how LangGraph interrupts and checkpointers let a person approve a step mid-run, plus the idempotency mistake that made my content agent publish something I never signed off on.
Claude Code forgets everything when the session ends. The setup I use: a post-mortem command that writes to CLAUDE.md, skill files for the quirks of the stack, and a daily pass over the logs.
Agentic loops fail quietly: the same query retried with new wording, retries with no circuit breaker, a context window full of raw JSON. Three patterns, how to measure them, and a PHP budget gatekeeper.
Offline evaluation for LLM agents: routing accuracy, LLM-as-judge scoring, and RAG faithfulness, plus the quality gate I use to block a deploy when the scores come in under threshold.
The API bill is the usual reason teams look at their own hardware, and privacy is the better one. Which benchmarks to trust for agents, how far you can quantize before logic breaks, what a single A100 on GCP costs, and serving Qwen 3.5-27B with vLLM.
Vector search handles synonyms well and loses exact identifiers. Notes on pairing BM25 with vector retrieval, merging both with rank fusion, and letting the agent retry when the results come back empty.