Short how-to guides and long-form deep dives on agentic engineering - architecture breakdowns and hard numbers, no filler. Reference cheatsheets for quick lookup live at Cheatsheets.
[ARTICLES: 42]Among 1,488 US workers, high AI oversight tracked +14% mental effort and +19% overload; ParallelPilot's PILOT practices and a two- or three-session ceiling cut the juggling tax.
Microsoft Execution Containers 1.0 enforces one JSON sandbox policy outside Codex and Copilot. Its denial log can draft that policy, but only on Windows, and --audit turns containment off.
Since Claude Code 2.1.292, stdio MCP servers get protocol 2026-07-28 by default. If tools vanish with 'Failed to fetch tools: Request timed out', read the MCP log, set MCP_PROTOCOL_NEGOTIATION=legacy, then fix the server or wait for the client fix.
On 23 September 2026 Cursor's harness changes cut user token cost 7%, including a system prompt about 66% shorter, rare tools loaded on demand, and line numbers on every tenth line.
POST /v1/decisions on gpt-6-luna costs $0.10 per million input tokens with no cache-read, cache-write, or output charge, and one 1,000-trial test saw a choice question return heads 98% of the time on a 70% coin.
"Compacted while idle, before the prompt cache expired" means Claude Code summarized a 200k+ session you walked away from. Set "idleCompaction": false (v2.1.290+) or take control with the API's compaction parameter.
Opus 5.5 and Sonnet 5.5 break old request code: disabled thinking, forced tool_choice and the old computer-use tool now return a 400. Per-model checklist.
Anthropic's 5.5 guides name four prompt habits that now backfire. /doctor prompt-audit found none in my repo - it found subagents for another workspace.
I ran claude plugin eval on obra/superpowers vs no plugin: it changed how Claude worked on 3 of 7 tasks, improved the result on 1, and cost 1.33x per task.
OpenAI's own guidance reduces to four moves: structure the ask, show one example, pick the answer mode, and correct instead of restarting. With a template.
Nobody can tell you if delegating to an agent will be faster. You can always know what it costs if it's wrong. Six rules for routing work by reversibility.
A rule in Claude Code's auto memory held 3/3 runs until the user contradicted it, then failed 0/3 - the same score the identical rule got in CLAUDE.md.
On the adversarial task its line-36 rule existed to prevent, a 129-instruction CLAUDE.md gave the same one-word reply as a harness with no instructions.
Anthropic's minimum cacheable prefix runs backwards: newer models have the lower floor, so one prefix caches on Opus 5 and silently never on Haiku 4.5.
Opus 5 keeps the $5/$25 price of Opus 4.8 with near-Fable agentic quality - within 0.5% of Fable on CursorBench at half the cost per task.
A Forward Deployed Engineer embeds in the customer environment to make frontier AI work on real data, workflows, and compliance - not from a ticket queue.
DeepSeek V4 Flash scores 79.0 on SWE-bench Verified vs Opus 5's reported 96.0. Route it as a cheap executor behind a frontier planner, not as the planner.
Every model on opencode's Zen free list is time-limited; January's free set had turned over by July. Adopt it for the harness and budget the paid fallback.
A high accuracy score means the tool passed one practice test - not that it can prove who wrote your essay. New models and 'humanizer' rewrites break it.
Decide open-weights vs closed API by data path, TCO, and ops burden, then hybridize by layer. Most 'open' models ship weights only, not training code.
Fast mode buys up to 2.5x speed on Opus for a flat ~2x per-token premium with no quality change; lower effort cuts tokens instead. Two controls that stack.
Prompt caching only pays when the prefix stays byte-identical - put stable content first, dynamic content last, and verify with cache_read_input_tokens.
Pick 1–2 custom LLM-judge metrics plus 2–3 generic agent metrics - about five. BLEU/ROUGE miss semantics; G-Eval, DAG, and QAG cover what humans grade.
Scope LLM-assisted builds with measurable goals, risk-first order, and tasks sized to one review sitting - plus a sunk cost checkpoint and a ready prompt.
Output tokens cost ~5× input and never cache - cut free-form prose first. Rewriting inside a cached prefix mid-session can raise the bill, not lower it.
A minimal Claude Code harness in ten files: a lean CLAUDE.md, a definition-of-done loop, read-only subagents, and credential deny rules. One npx install.
OS-level Bash sandboxing cut Anthropic's permission prompts ~84%, but /sandbox only wraps Bash. Pick the right isolation tier and deny credential reads.
Managing what a coding agent keeps in context - not window size - cut agent cost over 50% on SWE-bench Verified with no solve-rate drop, per JetBrains.
Prompt caching cuts 90–97.5% off repeated tokens, batching cuts 50%, and the two stacked cut 95% or more. A 2026 PwC study measured 41–80% real savings from caching alone.
Everything in one place: mindset, planning, the session loop, ten practices ranked by blast radius, playbooks, ten tools with 2026 pricing, and recovery.
Channels are MCP servers that push Telegram, Discord, iMessage, or webhook events into your open Claude Code session. Needs v2.1.80+ and an allowlist.
A practical map of Claude Code’s shortcuts, prefixes, session commands, and CLI flags - ordered like a real terminal cheatsheet.
Google's 2026 SDLC whitepaper draws the line between vibe coding and agentic engineering at verification, not AI use - and the harness decides your side.
Claude Code ships 46 built-in tools and only 14 ask for permission: the read/mutate split, daily gotchas, and ToolName(specifier) rules to restrict them.
Fable 5 is GA at $10/$50 per million tokens with a 1M context - and refusals arrive as HTTP 200, so integrations need refusal handling and a fallback.
/goal sets a completion condition; Claude keeps going until a separate fast model (Haiku by default) confirms it from the transcript. Needs v2.1.139+.
Plan mode makes Claude Code read and plan without editing until you approve (Shift+Tab). It blocks the edit tools - but shell commands can still write.
Claude Code runs a gather→act→verify loop that fills context before you type. Compaction, /clear, and subagents are the official levers for what sticks.
Fable 5.1 for the hardest reasoning, Opus 5.5 to orchestrate at $4/$20, Sonnet 5.5 as the default worker, Haiku 4.5 for parallel grunt work. Pick by job.
Sonnet 5 is the default on Free and Pro at $2/$10 per million tokens - close to Opus 4.8 on Anthropic's launch benchmarks at 40% of the Opus input price.
Claude Code runs up to a 1M-token window, but a bigger window doesn't fix a cluttered one - every turn replays history and quality drops before the cap.
A bloated CLAUDE.md gets filtered out as noise. Keep it to universal instructions and move everything else behind progressive disclosure.