Behind every buzzword there is a mechanism.
The terms everyone throws around in AI engineering, taken apart: where it came from, what it actually does, and what we think after running it ourselves.
89 tasks gone: what happened to SWE-Bench Pro, and what to check in your eval
Scale AI pulled 89 tasks from SWE-Bench Pro after agents were caught reading answers from git history. What happened, and what to check in your eval.
Read →Series
The new tools, up close
- Codex up close: what OpenAI actually installs on your machine
- Grok Bot up close: a cloud computer that holds your passwords
- ChatGPT up close: the features engineers leave switched off
- Claude Design up close: the tool that reads your repo and draws you a deck
Claude Code in practice
- Claude Code in Practice, Part 1: Short CLAUDE.md, Strict Permissions
- Claude Code in practice, part 2: hooks, MCP and skills, and what goes where
- Claude Code in practice, part 3: subagents, worktrees and a clean context
- Claude Code in practice, part 4: in CI, exit 0 isn't success
- Claude Code in practice, part 5: which model, which mode, what it costs
The 2026 assistants, one by one
- OpenAI dots up close: the agent that never goes home
- Gemini up close: cheap, fast, and changing every three weeks
- Microsoft Copilot up close: a permissions project dressed up as AI
- Perplexity Comet up close: a great research assistant, and an agent you're better off not running
Prompt engineering, minus the magic
- Prompt engineering in 2026: what's still worth writing in the prompt
- System prompts in production: four blocks, one order, and a cache that saves 90%
- Prompt evals: how Rechat's Lucy stalled, and how to know a prompt change helped
More terms
53
Same engineering: same tasks, a different score, and no model changed
DeepSWE v1.1 changed only execution and grading, and tasks moved 36 points. How to measure this yourself without mistaking noise for progress.
Small coding: a code agent with a local model, a one-file task, and a loop that checks itself
How to build a small coding agent on Ollama: a task with a verify command, a 50-line harness, trimmed context, and when to hand off to a big model.
Agentic software: the DELETE no prompt could stop
Replit's agent deleted 1,206 records during a code freeze. What actually broke, why another prompt won't help, and what to check in your agent this week.
AI-native engineering: writing code isn't the bottleneck anymore, verifying it is
What AI-native engineering is, what Meta and Anthropic actually did, and why the work moved from writing code to verifying it. With an AGENTS.md to copy.
Since coding: five years after Copilot, and what to change on your team on Monday
The model is the least important call when working with coding agents. Six steps: permissions, a devbox per task, AGENTS.md, proof in the PR, measurement.
Top coding: the list that went stale in 106 days
On June 8 Opus 4.8 scored 88.6%. By September 22 Claude Code ran a different model. What happened to coding agent rankings, and what to pin this week.
Accelerating engineering: not a term, a workflow with gates
Accelerating engineering is a headline, not a method. What works at LinkedIn and AWS: a narrow workflow, human gates, and a check at the end.
Ends coding: the code isn't over, the typing is
The agent writes most of the lines. The engineering moved around it: spec, sandbox, deny in a file, test first, diff review. Six steps.
38% with no internet: what breaks when you run a coding agent on a local model
Antigma ran 8 open models on 89 Terminal-Bench tasks with no network. The 27B hit 38%; the 4B broke on tool calls. Here's what to check.
Agentic workflows: who picks the next step
What an agentic workflow is, how it differs from an agent, what GitHub, Elastic and Pyrefly learned, and why a 30-second keepalive wastes money.
Agentic operating: how to run ops with agents without buying an operating system
The agentic operating system isn't a product. It's a runbook a model runs. Six steps to run one loop with agents, from permissions to measurement.
Respective coding: Replit's agent, 11 instructions and one deleted DB
In July 2025 Replit's agent wiped a production DB during a code freeze. What went wrong, why more prompt rules don't help, and what to check this week.
Spec-driven development: you edit the spec, not the code
What spec-driven development is, why it isn't waterfall, how spec, design and tasks work, and how one spec line yields two documents with the same name
Advancing coding: migrating COBOL with an agent without losing a cent
COBOL migration with an agent in six steps: golden master, an AGENTS.md contract, byte-level diff, and high reasoning effort only for what failed.
by coding: the Neon error that moved from boot to the first query
An agent reviewing Neon found 1.2.0 moved the missing-env error from boot to the first query. What that says about code agents write.
Agentic AI: who picks the next step
Agentic isn't a model or a product. It's one question about your system, and how to check it on Monday: Ng, Anthropic, Replit and a checklist.
Agentic AI isn't a model, it's a permission. Six steps before an agent touches anything
Agentic AI starts when the model both picks the next step and runs it. Six practical steps: identity, a DB role, a checker in the loop, a human gate.
11 times CODE FREEZE: what the Replit deletion teaches people who write skills
In July 2025 a Replit agent deleted a production DB despite 11 code-freeze instructions. What that means for skills, and what to move from text to code.
Graph engineering: when one loop stops being enough
Graph engineering without the hype: why agent loops get stuck, how to turn an invariant into a LangGraph edge, and when a plain loop is still the better call.
Agent-native architecture: count the buttons before you build the agent
A parity map, one registry for UI and agent, and a gate on irreversible actions. Six steps to make an existing product agent-native, with code to copy.
Context engineering: how Anthropic cut 80% of Claude Code's prompt and lost nothing
Anthropic cut over 80% of Claude Code's system prompt with no drop in evals. What went wrong in the context, and how to audit your CLAUDE.md this week.
Loop engineering: who decides the loop is done
What loop engineering is, where it came from in June 2026, where it breaks, and a minimal bash loop with a check, a budget and memory you can run tomorrow.
Forward Deployed Engineering: six steps to put an agent into someone else's org
620 open FDE roles, $4B for OpenAI's DeployCo. What it takes to put an agent into a customer's org: six steps, with evals and shadow mode.
MCP Security: how one config field became RCE across 150 million downloads
In April 2026 one MCP STDIO design flaw produced 10 CVEs across 150M downloads. What broke, why the filter failed, and what to check this week.
Agent Skills: your agent doesn't need another doc, it needs a workflow
Agent Skills explained: what loads into context and when, what SkillsBench measured, and a SKILL.md you can copy this week.
Vibe engineering in six steps: the model writes, you decide
Vibe engineering in practice: a spec with open decisions, AGENTS.md, plan mode, a hook that enforces tests, and a diff review. Six steps, files to copy.
1,206 records and one code freeze: what Replit's agent teaches about agentic engineering
In July 2025 Replit's agent deleted SaaStr's database mid code freeze. Where it failed, why another prompt wouldn't help, and what to check this week.
Agentic evals: check what the agent did, not what it wrote
Agentic evals: why grading the final text misses agent failures, three layers to check, pass^k, and a pytest file you can copy today.
GraphRAG: three Postgres tables before you buy a graph DB
GraphRAG without a graph DB, in six steps: a 20-question eval, an ontology, three Postgres tables, and a single retrieval query.
Prompt engineering: six words and a command with no due date
Michael Heap gave an agent six words and got a command without a due date. What went wrong, why regenerate doesn't help, and what goes in TASK.md.
Long-running agents: how an agent works for 36 hours without losing the thread
Long-running agents without the hype: why 1M tokens isn't enough, a loop with a progress file and automated checks, and a one-page harness to copy.
Agentic skill: how to write a SKILL.md that doesn't become another system prompt
Six steps to an Agent Skill that works: a description that triggers right, a body of steps, a checking script, a 10-request eval and an expiry date.
Agentic memory: why an agent with 200K tokens forgot how to write a date
An agent on a Raspberry Pi with 200K tokens of context forgot a one-line rule. What broke, why a vector DB didn't help, and what to check.
Agentic RAG in practice: six steps from one search to a capped loop
How to turn one-search RAG into an agent that searches, checks and searches again: eval, three tools, a grader, MAX_STEPS and a router, with code.
3/4ChatGPT up close: the features engineers leave switched off
Projects, memory, canvas, deep research, agent and connectors in ChatGPT: what each one does, where it breaks, and what to turn on this week.
1/5Claude Code in Practice, Part 1: Short CLAUDE.md, Strict Permissions
The three CLAUDE.md files, auto memory, settings.json permissions, and the commands worth memorizing. Step by step, for a real team.
2/5Claude Code in practice, part 2: hooks, MCP and skills, and what goes where
CLAUDE.md says what's required, a skill says how, MCP what's possible, a hook enforces. A PreToolUse block on main, auto-formatting, MCP, a first SKILL.md.
3/5Claude Code in practice, part 3: subagents, worktrees and a clean context
When a subagent saves context and when it burns 15x the tokens, a git worktree per parallel session, plan mode, and a three-line handoff.
4/5Claude Code in practice, part 4: in CI, exit 0 isn't success
claude-code-action, claude -p with JSON, and the Agent SDK: a workflow to copy, subscription vs API, cost caps, and five checks after every run.
5/5Claude Code in practice, part 5: which model, which mode, what it costs
Sonnet, Opus or Haiku per task, effort and fast mode, four permission modes, AGENTS.md vs CLAUDE.md, and reading /usage. A command and a number for each.
4/4Claude Design up close: the tool that reads your repo and draws you a deck
Anthropic's Claude Design: what it builds, how it learns your design system from code, what it costs, where it fails, and when Open Design or Figma win.
1/4Codex up close: what OpenAI actually installs on your machine
OpenAI Codex in late 2026: CLI, cloud, IDE and mobile. What it does well, what it costs, and how it compares to Claude Code day to day.
2/4Gemini up close: cheap, fast, and changing every three weeks
Gemini in late 2026: CLI, Live, Workspace and Flash pricing. Where it beats ChatGPT and Claude, where it falls short, and when an Israeli team should pick it.
2/4Grok Bot up close: a cloud computer that holds your passwords
SpaceXAI's Grok Bot: what it does on its own, what it costs when sources disagree, where it breaks, and the open-source alternative.
Harness engineering: what broke when OpenAI let Codex write a million lines
OpenAI built a product with zero hand-written code. The code drifted, the first fix failed, and harness engineering came out of it. What to check in yours.
Jev: when your agent needs a decision, not a paragraph
TypeSafe's Jev returns a decision with a probability instead of text. What it actually does, what it costs, and how to build a decision layer this week.
3/4Microsoft Copilot up close: a permissions project dressed up as AI
Copilot in 2026: four products, per-seat pricing, what works in a Microsoft shop, where it leaks, and a rollout checklist for 50 people.
1/4OpenAI dots up close: the agent that never goes home
OpenAI dots up close: an agent on its own cloud computer, running 24/7 across 4,000 apps. What it does alone, what it costs, and what not to give it.
4/4Perplexity Comet up close: a great research assistant, and an agent you're better off not running
Perplexity Comet up close: it reads and compares pages very well, and its agent works inside your session. Pricing, risks, and five research prompts.
1/3Prompt engineering in 2026: what's still worth writing in the prompt
Models understand plain language, and most 2023 tricks are dead. What moved into context and the harness, and what still belongs in the prompt.
2/3System prompts in production: four blocks, one order, and a cache that saves 90%
How to build a production system prompt: role, rules, examples, output contract, XML vs markdown, few-shot, length, and prompt caching with real prices.
3/3Prompt evals: how Rechat's Lucy stalled, and how to know a prompt change helped
Part 3: golden sets, LLM-as-judge, pairwise comparison and regression on real traffic. The Lucy case, plus a promptfoo config to paste.