Deep dives into working agents.

Technical walkthroughs with substance: architectures, prompting patterns, evals, and production lessons unpacked until complex agent systems become reusable design material.

Page 2 of 4 · showing 25–48 of 78

Show the outdated contentHistory

Are Agent Swarms Useful? OpenAI's GPT-6 Astra Swarm Takeaways

OpenAI got hacked by their own agents — the GPT-6 Astra swarm incident, where isolated EVAL agents built their own message board in a package cache, had it wiped, and rebuilt it. IndyDevDan turns the incident into engineering action: build a V1 swarm system and run three real multi-agent experiments on an isolated M4 Mac mini (GLM 5.3 10-agent pelican for $20/56min, DeepSeek v4 Pro 20-agent ray tracer, Gemini 3.7 Flash 30-agent canvas recreation). Three takeaways: communication is the unlock (give every agent a mailbox/thread); alignment + definition of done (OpenAI gave impossible tasks with no bail-out — encode a done tool as a harness rule); observability + sandboxing (watch every tool call, isolate the swarm, don't run it on your daily driver). Swarms sit between software factories and dark factories — and it's the first tech he won't open-source.

deep-dives

IndyDevDan

Published2026-09-07

Claude Code Just Made CLAUDE.md Feel Obsolete

Anthropic released function hooks — which Ray Amjad calls the best feature in Claude Code yet. They upgrade shell hooks from simple pre/post scripts into Express.js-style middleware: on any tool call you can block, rewrite input, short-circuit, append context, draw custom UI rows, ask the user questions, add tools, and persist memory across hooks and sessions. Two flagship demos: a secret redactor that swaps API keys for in-memory IDs so secrets never touch the transcript, and a live Vercel deploy status row. Plus a catalog of patterns: dry runs, a >1000-line refactor gate, turn-end TTS summaries via a Haiku model, compliance audit trails, newsletter send-guards, knowledge-base injection, and PR quizzes. The thesis: migrate the rules in your CLAUDE.md into deterministic, shareable hooks.json plugins.

deep-dives

Ray Amjad

Published2026-09-04

11 Tiny Coding Agent Fixes With a Stupid Amount of Payoff

Cole Medin's 11 easy, agent-agnostic tips to make coding agents more reliable, each backed by a study: write for the agent (specific paths/commands, not human prose); your instruction files rot (1 in 4 repos have stale rules — audit with rules-check-drift); /compact is not worth it (only ~10% of detail survives); put load-bearing rules in hooks (deterministic guarantees, not probabilistic rules); less context is more (keep global rules <200-300 lines); stop liberal sub-agents (39% of a weekly limit went to parallel sessions); don't escalate mid-task (write a handoff doc and burn the conversation); skip the coordinator/team-lead (delegation beats agent-teams); never let the writer approve its own work; you can over-revise (85% had an earlier iteration better than the last); and treat validation as a system planned before code.

deep-dives

Cole Medin

Published2026-09-01

Agentic Engineering Operating Level — Where to Focus Your Agents

IndyDevDan's framework for where you and your agents should spend time, focus, and attention. Five operating levels — Code Primitives, Code Structure, Data and Execution, Delivery and Intent, Agentic Systems — where moving up gains leverage and speed, and moving down gains control and understanding ('you do not get both'). The core theses: higher is NOT better (leverage on top of ignorance is a trap); range beats position (the difference between a vibe coder and an agentic engineer is how many levels you can operate on); 'three makes a pattern' (the trigger to automate); the data level matters (own your database tables); and out-of-distribution isn't just what the model doesn't know but what it actively fights you on. Closes with the ladder beyond the software factory: dark factory and RSI.

deep-dives

IndyDevDan

Published2026-08-31

WikiSkill: Evolving Skills for LLM Agents

AI Research Roundup on Google Research's WikiSkill (Aug 27, 2026). The core problem: valuable agent lessons are lost over time because error logs scatter across messy run records. WikiSkill co-evolves skills with a persistent knowledge base across three tiers (raw traces / persistent wiki / executable skills), with a validation-gated rollback that keeps failed knowledge. Results: highest average accuracy across five models, beating the strongest baseline (EvoSkill, SkillOpt) by 3-12 points; smaller models with evolved skills beat larger baselines without skills; skills transfer across model families (Qwen 27B's skills boosted Qwen 9B from 34% to 50%+ on spreadsheet); and the ablation shows the wiki must live in the skill proposer's loop.

deep-dives

AI Research Roundup

Published2026-08-30

Claude Code Got INSANE and MASSIVE Updates (Week 32, Aug 2026)

Anthropic's Week 32 (August 2026) Claude Code updates: cross-session messaging (v2.1224) lets terminals communicate with concise, proactive updates (/list agents; macOS/Linux); auto mode is now default (a background classifier caught 89% of dangerous actions vs humans' 13.6%, zero prompt-injection successes in 720 attempts, +25% PR velocity); concise output mode, 2x launch speed, built-in iOS simulator (beta), screen-reader mode, and Ultra /code review (multi-agent, fewer false positives); plus MCP-powered artifacts and a sandboxed built-in browser. Pro tip: pair with Opus 5.

deep-dives

JustAIWorld

Published2026-08-29

37 Cheat Codes to Level Up in Claude Code

Simon Scrapes catalogs 37 Claude Code tricks in six buckets: make Claude talk less (Simple English, Stop Slop, Caveman, concise output style), hidden keyboard shortcuts (Ctrl+R/S/T/G/O/P/A/W, Option+T, /focus), never losing a session (rewind vs fix, resume-from-summary, /rename, @ file targeting, claude-leport), token bill fixes ('give me a quick answer' to skip cache rebuilds, show-clear-context-on-plan-accept, repo CC usage, /recap), memory that pays rent (merged memory files, path-scoped claude.md rules, HTML comments, hand off as a prompt not a summary), and the newest tricks (5-deep skill nesting, 3-part skill descriptions, Claude Design over the front-end skill, /permissions mid-session, image annotation, memory import, scheduled tasks in project context).

deep-dives

Simon Scrapes

Published2026-08-28

Intelligence EXPLOSION: Harness Engineering with Pi Agent, DeepSeek & Gemini

IndyDevDan ships V2 of his Fusion Harness, a custom Pi coding agent for multi-model orchestration. Three patterns — Opinion (N models, one prompt), Debate (agents argue with hidden aliases), Collaborate (plans merged by an architect agent) — running Fable 5, Gemini 3.7 Flash, and DeepSeek V4 Pro side by side on a DuckDB v2.0 task.

deep-dives

IndyDevDan

Published2026-08-24

19 Claude Code Mistakes "Pro" Users Are Still Making

Simon Scrapes' 19 rookie mistakes even pro Claude Code users make — personas are dead, 'do not' instructions confuse the model, MCPs no longer eat context, subagents cost 7x for context-heavy work, fast mode is an API-credit trap, Claude.md should be ~300-350 words, and the 4-level verification ladder. Plus /doctor, /insights, /btw, /branch.

deep-dives

Simon Scrapes

Published2026-08-22

FIXING Opus 5: PROOF that Prompt Engineering IS NOT DEAD

IndyDevDan turns verbose 'smartass' Opus 5 into a precise senior engineer using system prompt engineering. Five repeatable techniques: positive+negative patterns, reference points (D/R/F/P codes), hard operational boundaries, prompt aliases (SCR/ELI/FOC/REF), and in-context-distillation examples. The system prompt is the law — and where real leverage lives.

deep-dives

IndyDevDan

Published2026-08-17

Andrew Ng's Graph Engineering Course — 1 Prompt → 100 Agents → Loops → Graphs

A 2-hour hands-on course with Andrew Ng, Harrison Chase (LangChain), and Doren Vice (Tavily): building agents from scratch with the React pattern, LangGraph's nodes/edges/state architecture, agentic search, human-in-the-loop, persistence with checkpoints, streaming, state modification, and time travel — all with running code.

deep-dives

Function Form (Andrew Ng, Harrison Chase, Doren Vice)

Published2026-08-10

Engineers… Your Software Factory NEEDS Agent Sandboxes to SCALE (exe.dev)

IndyDevDan puts the entire software factory inside an agent sandbox. The three-tier architecture (out-loop orchestrator, in-sandbox orchestrator, ADW agents), Best of N across five agent configs, disposable OpenRouter keys for a bounded blast radius, and why isolation+scale+autonomy beats a container. Agentic engineering vs vibe coding.

deep-dives

IndyDevDan

Published2026-08-10

Meta's Claude Code Clone is INSANELY Cheap — Muse Code & Muse Spark 1.2

Theo - t3.gg puts Meta's new Muse Code terminal agent through its paces: benchmarks against Opus 5 and GPT-5.6, a 10-20x pricing gap on the contributor tier, 191 TPS throughput, live codebase audits, head-to-head comparisons with Fable and DeepSeek, and the killer use case — 222 PRs audited in 5 minutes for 10 cents.

deep-dives

Theo - t3.gg

Published2026-08-07

Prime Agent: The Architecture That Makes Self-Improving Agents Safe — Full Breakdown

A comprehensive deep dive into Prime Agent's four design decisions from Signal Coders' source-code review: the one-tool IPython architecture, fire-and-forget sub-agents, the self-improvement system with its immutable base and snapshotted learning, and autonomous mode with four budgets and quality gates — plus the most honest sentence in agent documentation this year.

deep-dives

Signal Coders

Published2026-08-07

Eight Skills Every Agentic Engineer Needs: David Ondrej's 3.4K-Star Toolkit

A thirty-minute tour of an open-source skills repository that went viral — 42 skills across five categories, 3,418 stars. David Ondrej walks through his eight favourites: global agent guardrails (pre-tool-call hooks), git worktrees, VPS management, Go loop, decisions, anti-sleep, a research agent and the web search skill used hundreds of times per day. Endorsed by the Hermes co-founder and OpenSea CTO.

deep-dives

David Ondrej

Published2026-08-05

You Pay Twice: Aggregate Usage Data and the AI Sovereignty Ladder

IndyDevDan answers the headline question in the first third — no, Anthropic is not stealing your data — and spends the rest on the harder problem underneath. A reading of the actual terms of service, a four-vertical pattern of platforms entering their customers' markets, the commodity-vs-IP agent test, and a six-tier sovereignty ladder for deciding how much of your stack you actually own.

deep-dives

Editorial

Published2026-07-27

TypeScript 7 Is Here (And It's 10× Faster)

Anders Hejlsberg announces TypeScript 7 — the Go port of the compiler and tooling that delivers 10× speed improvements through native code and shared-memory concurrency, compiling VS Code's 1.3M lines in under 5 seconds with a new LSP language server.

deep-dives

Microsoft Developer

Published2026-07-24

I Created the Ultimate Coding Agent by Combining Pi, Codex, and Claude Code

Ben Davis builds a unified orchestration layer in Pi that spawns sub-agents across Pi, Codex, and Claude Code — leveraging GPT-5.6 for speed, Fable for code quality, and Codex for computer use, all from one terminal.

deep-dives

Ben Davis

Published2026-07-24

Build an Agent with Eve — The Open Agent Framework

Andrew and Casey, tech leads on Eve at Vercel, live-code a declarative agent from scratch at Ship 26 NYC — scaffolding with npx init, deploying to Slack via Vercel Connect, adding OpenAPI connections, MongoDB tools with human-in-the-loop approvals, evals, and the stateless Workflow SDK architecture that lets agents run for hours.

deep-dives

Vercel

Published2026-07-20

pgrust: What Changed in the Thirteen Days After This Video

A Rust reimplementation of Postgres passing all 46,066 official regression queries is real and verifiable. But the video's central caveat - that the interesting version is unpublished and unbenchmarkable - expired within two weeks. The shipped v0.2 revises the performance claims downward (50% becomes 30%, with the maintainers explaining why they quote the lower figure), and buries a finding more interesting than any speed number: formal verification of 1,000 functions surfaced four bugs in Postgres itself.

deep-dives

Better Stack

Published2026-07-17

System Prompts Leaks GitHub Walkthrough: ChatGPT, Claude, Gemini & Guardrail Analysis

Walkthrough of the asgeirtj/system_prompts_leaks repo (58K+ stars, CC0). Documents hidden system prompts from every major AI: Claude, ChatGPT, Gemini, Grok, Copilot, Perplexity, Cursor, and more. Reveals how each company configures guardrails, personalities (GPT-5.1 has 8 modes), and tool access (Claude Design: 48 tools, 16 skills). DiffChecker links track guardrail evolution between versions.

deep-dives

Alex Hitt

Published2026-07-16

Claude Code's System Tools Are SO BLOATED — Matt Pocock

Matt Pocock shows how Claude Code's default system prompt ships at ~25K tokens of mostly unused tool definitions. By disabling unused features in settings.json (plan mode, crons, bundled skills, workflows, remote control, AI connectors, artifacts), he cut it to 8K tokens — a 68% reduction that improves output quality.

deep-dives

Matt Pocock

Published2026-07-16

AI Hacking Got Too Easy — Free Tools Beat a $1B Bot

XBOW topped HackerOne's US leaderboard at $1B valuation — then free open-source tools (Shannon, Strix, Tempest) matched it on its own benchmark. MCP + HexStrike plugs 150+ hacking tools into any AI. The price of autonomous hacking fell from $6,000 to $0.

deep-dives

Bitwise AI

Published2026-07-11

The Ultimate Claude Code Tutorial for Mobile Apps — FULL COURSE

A 3.5-hour complete build of 'Triply', an AI trip planner (React Native + Expo, Clerk, Neon, Drizzle, Inngest, Sentry, ImageKit, CodeRabbit). The extended deep dive documents the full structured workflow: plan-mode interview, GPT-image UI design + the 9-variation trick, the screenshot-compare verify loop, agents.md vs claude.md, Clerk skills, webhooks (Clerk→Neon user sync), native tabs (liquid glass), trip generation with Unsplash+ImageKit, Sentry logs/traces/replays/AI-agent monitoring, Inngest experiments (90/10 model routing), and the legal pages + Cloudflare deploy — plus the complete troubleshooting log and App Store requirements.

deep-dives

Codesistency

Published2026-07-04