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: 33]OpenAI's own guidance reduces to four moves: structure the ask, show one example, pick the right answer mode, and correct instead of restarting. With a copy-paste template.
Nobody can tell you in advance whether delegating to an agent will be faster. You can always know what it costs if it is wrong. Six rules for routing decisions by reversibility.
One rule, three layers, measured: 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 one adversarial task its line-36 rule existed to prevent, a 129-instruction CLAUDE.md returned the same one-word reply as a harness with no instructions at all.
Anthropic's minimum cacheable prefix is model-dependent and runs backwards - newer, pricier models have the lower floor, so the same prefix can cache on Opus 5 and silently never cache on Haiku 4.5.
Opus 5 ships at the same $5/$25 as Opus 4.8 with near-Fable agentic quality - more than double Opus 4.8 on Frontier-Bench, within 0.5% of Fable on CursorBench at half the cost per task - so escalate to Fable only when the peak or the refusal surface is worth it.
A Forward Deployed Engineer embeds inside the customer environment to make frontier AI work against real data, workflows, and compliance - scope, eval, deliver - not from a ticket queue.
Flash scores 79.0 on SWE-bench Verified against Claude Opus 5's 96.0, and 60/100 on a real agentic build. Route it as the cheap executor behind a frontier planner, at 1/89th of Opus 5's output cost.
Every model on opencode's Zen free list is time-limited, and January's free set had almost entirely turned over by July. Adopt it for the harness, and budget the paid fallback before it arrives.
A high accuracy score often means the tool did well on one practice test - not that it can prove who wrote your essay. Wrong accusations, new AI models, and 'humanizer' rewrites all punch holes in that number.
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 or data.
Fast mode buys up to 2.5x speed on Opus for a flat $10/$50-per-million-token premium with zero quality change, while a lower effort level cuts tokens - and sometimes quality - instead; they're independent controls meant to stack, not substitute for each other.
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 for the use case plus 2–3 generic agent metrics - about five total. BLEU/ROUGE miss semantics; G-Eval, DAG, and QAG cover what humans actually grade.
A scoping method for LLM-assisted builds: measurable goals, risk-first order, and task size capped by what you can review in one sitting instead of by the calendar - plus a sunk cost checkpoint and a ready-to-copy planning 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 fits in ten files - a lean CLAUDE.md skeleton, a verifiable definition-of-done loop, three read-only subagents, a pre-ship security audit, and credential deny rules. One npx command installs the set.
OS-level Bash sandboxing cut Anthropic's permission prompts ~84%, but /sandbox only wraps Bash. Pick the right isolation tier, deny credential reads, and lock managed settings for unattended work.
Managing what a coding agent keeps in context - not window size - cut agent cost over 50% on SWE-bench Verified with no drop in solve rate, per JetBrains Research.
Prompt caching cuts 90% off repeated tokens and batching cuts 50% flat - 95% stacked. A 2026 PwC study measured 41–80% real savings from caching alone on agent workloads.
Everything in one place: mindset, planning, the session loop, a worked example, ten practices ranked by blast radius, git/prompting/debugging/testing/security playbooks, ten tools with 2026 pricing, and the commands that recover a session gone wrong.
Channels are MCP servers that inject Telegram, Discord, iMessage, or webhook events into the local Claude Code session you already have open - v2.1.80+, `--channels` per session, and a sender allowlist, not just `.mcp.json`.
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 around the model decides which side you're on.
Claude Code ships 42 built-in tools and only 13 ask for permission - the read/mutate split, gotchas for the tools you touch daily, and the ToolName(specifier) rules that restrict or disable any of them.
Fable 5 ships GA at $10/$50 per million tokens with a 1M context - and safety classifiers that return refusals as HTTP 200 responses, so every integration needs refusal handling, a fallback path, and new billing rules.
/goal sets a completion condition and Claude keeps taking turns until a separate small fast model (default Haiku) confirms it from the transcript. Requires v2.1.139+, one goal per session, condition up to 4,000 characters.
Plan mode makes Claude Code read and plan without touching disk until you approve - enable it with Shift+Tab or --permission-mode plan. Under the hood it's prompt reinforcement plus a markdown plan file, not a hard tool lockdown.
Claude Code runs a gather→act→verify loop: startup load fills context before you type, then file reads and tool output pile up as the loop runs - compaction, /clear, and subagents are the official levers for what sticks.
Fable 5 escalates for the hardest reasoning, Opus 5 orchestrates near Fable at half the price, Sonnet 5 is the default worker, Haiku 4.5 does the parallel grunt work. Choose by the job and the token bill, not the benchmark.
Sonnet 5 ships as the default on Free and Pro with intro API pricing at $2/$10 per million tokens - near Opus 5 on agentic benchmarks at 40% of Opus input cost through the 2026-08-31 intro window (60% at standard pricing) when you keep effort at Medium or below.
Claude Code runs up to a 1M-token window on flagship models, but a bigger window doesn't fix a cluttered one - every turn replays the whole 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.