⚙
Reading for Engineering
Building with agents — specs, orchestration, evals, cost.
Agentic Engineering12 posts
The flagship series.
- 01 · The Review Is the Work Now Developers fear agentic AI is pulling them out of the code and that reviewing it eats their time. The real shift isn't coding-plus-overhead — it's soloist to conductor, Organic Intelligence to AI.
- 02 · The Same Way Twice An agent that does the same task two different ways hasn't given you leverage — it's given you a thing you now have to re-review every run. Spec-driven development is how you get determinism, and determinism is the floor that runtime governance is built on.
- 03 · Easy Bake Oven Developers There's a kind of developer who treats AI like a toy oven — toss in random ingredients, expect a finished cake, and blame the appliance when it doesn't work. The problem was never the model. It's GIGO, and it's a mental-model problem.
- 04 · Intelligence Moves Up the Stack Agentic development isn't a chat window — it's a cognitive abstraction on top of one. Control just inverted again, all the way up to intent. The people who thrive are the ones who can turn intent into concrete, deterministic output.
- 05 · Anatomy of a Spec The series kept insisting you need a spec. None of it told you how to write one. Here's the anatomy — the parts most specs are missing, the altitude problem that's the actual craft, and the test that tells you it's done.
- 06 · Jerry Maguire These Agents: SHOW. ME. THE. RECEIPT. Agents will confidently tell you they did the thing — when they didn't. The fix is a trust pyramid: push every claim from self-reported up to evidence-backed and independently verified. Don't believe anything an agent says about reality until it shows you the receipt.
- 07 · Orchestration A spec you write and walk away from is a New Year's resolution. The hard part isn't defining intent — it's keeping it in force: guardrails that hold the line, a workflow that drives the spec, a plan that outlives the session, and drift detection that catches the gap.
- 08 · Garbage In, Gospel Out: Agentic Cognitive Bias Humans have confirmation bias — we read new evidence as support for the hypothesis we already hold. Agents have it worse, because they can't tell where a 'fact' in their context came from. Once something false enters the window, every next token treats it as gospel. That's context poisoning, and it's the failure mode underneath most agent train wrecks.
- 09 · When the Stack Pushes Back This whole series argues that your value moves up — to intent, to orchestration. This is the honest counterweight: the places where climbing the stack costs more than it pays, and how to tell you're standing on one.
- 10 · The Team Dimension Everything in this series so far has been about you — your judgment, your intent, your review. But moving up the stack is a team sport, and a team that climbs unevenly doesn't just move slower. It rots while shipping more.
- 11 · Multi-Agent / Distributed Intent Directing one agent is orchestration. Directing many is composition — and the moment your intent fans out across agents, behavior nobody specified starts showing up in the gaps between them. Here's how to compose intent without losing the plot.
- 12 · Intent-Driven Development I've written a series on how to spec and a series on how to eval. This is the post where I admit they were never two skills. A spec is your intent declared; an eval is the contract that proves it. Do both on purpose and you have a method — Intent-Driven Development.
Leadership in the Agentic Era7 posts
Leading teams + orgs through the shift.
- 01 · Acquiring Talent in the Age of AI The interview became an open-book exam and everyone brought the book. Companies are responding by trying to catch the AI — banning tools, dragging candidates back on-site. They're solving the wrong problem. The job changed; the test didn't.
- 02 · Leveraged Time You learned to code by getting fast at one instrument, heads-down, single-threaded. That skill is now the trap. The unlock is putting the violin down and picking up the baton — spreading your will across a team of agents instead of spending your hours.
- 03 · Lean on the Tools Conduct ten agents and you stop being the soloist — you become the bottleneck. Not your typing. Your review capacity. The fix isn't reviewing harder; it's letting specs, evals, and tests take the first pass so your judgment lands only where it's needed.
- 04 · Measure the Right Thing Story points, velocity, tickets closed — those instruments were built to measure the one thing that just got cheap. In an agentic world they don't just miss the point. They actively point the wrong way.
- 05 · The Incongruent Org A team doesn't climb the stack all at once. When half of it conducts agents while half still plays solo — and dev races ahead of product — the org doesn't just slow down. It tears along the seam where intent runs out.
- 06 · Leading the Change You can't org-chart your way through this. Leading the agentic shift is empathy first, foundation second, language third — in that exact order. Skip the first and you lose people; skip the second and you scale chaos; skip the third and no one can reason together.
- 07 · Mandate, Then Get Out of the Way You can't tell people to trust these tools — trust is earned by using them, and the payoff is back-loaded, so people quit before the reps land. The fix is a forcing function: mandate the reps, coach hard, then take the whip away.
The New Ways of Working7 posts
How the job itself is changing.
- 01 · The New Caste System: Who's Drowning in Agentic Dev — and Where the Life Rafts Are A new caste system is forming around agentic AI. It has nothing to do with talent and everything to do with fluency. Here is the ladder — and how to climb it before the cheap-token party ends.
- 02 · Upgrade Your Workflow Working with agents isn't a faster version of the old job — it's a different operating system. Here are the six meta-skills that run on it, and the series that unpacks each one.
- 03 · Voice-Driven Dictation You can talk roughly three times faster than you type, and your agents don't care which one you use. Here's why voice is a first-class input to an agentic workflow — and where it's still genuinely clumsy.
- 04 · Loops The one-shot prompt is the training wheels of agentic work. The real leverage is the standing loop — a job that runs on an interval or runs until a condition is met, working while you don't.
- 05 · Code the Interface, Delegate the Implementation The interface is intent made precise and verifiable. The implementation is the cheap, fungible part now. Here's the working pattern: you write the contract, the agent fills in the body.
- 06 · When Agents Take Over the Office The migration from organic intelligence to AI feels like a threat because we keep framing it as replacement. It's a promotion — here's how to make the transition painless and free up the part of your brain that's worth paying for.
- 07 · Ambition → Intention → Execution Every built thing travels the same path: from wanting it, to specifying it, to making it. Agentic development didn't change the path — it changed which step is scarce. Execution got cheap. Intention became the whole job.
Token Economics5 posts
What your agents cost, and how to trim it.
- 01 · Measuring Agent Spend Over Time You can't manage what you can't see. Before you optimize a single token, you need to know where the money goes — by session, by user, by agent, by model. Here's how to actually measure it.
- 02 · Analyzing Token Trends, Regardless of the Harness A single month's token bill tells you almost nothing. The number that changes decisions is the one you can watch move over time — normalized across every harness, attributed to features and people, tied back to value.
- 03 · Trimming Token Cost Most teams overpay for agents by default — re-sending the same context, running a flagship model on trivial work, and stuffing the window with junk. Here are the levers that cut the bill without cutting quality.
- 04 · Routing: LiteLLM, OpenRouter, and Automatic Token Routing A routing layer isn't just a cost lever — it's a resilience and independence lever. One API across every provider, easy tasks to cheap models and hard ones to strong models, and failover when a lab falls over.
- 05 · Caching: Prompt, LLM, and Semantic Agents re-read the same context on every turn and re-answer the same questions all day. Three kinds of caching kill that waste — but each one has a failure mode that quietly serves you the wrong answer.
Ship Confidence5 posts
Evals — ship on proof, not vibes.
- 01 · What Are Evals? The gap between 'it worked when I tried it' and engineering has a name: evals. They're repeatable, measured tests of an AI system's output quality — and they're the difference between shipping on confidence and shipping on hope.
- 02 · An Intro to DeepEval You decided evals matter. Now you need to actually write one. Here's DeepEval from zero — what a test case is, how metrics judge it, and how to read your first red bar.
- 03 · A Custom G-Eval Grade with a Claude Subagent Off-the-shelf metrics never fit your domain. Here's the concrete build: a Claude Code subagent that acts as a G-Eval judge, wired into Vitest so a rubric you wrote becomes a pass/fail in CI — with the full agent definition and the test code.
- 04 · How Evals Enable Self-Improving Harnesses Evals don't just catch regressions — they turn a harness into a system that gets better on purpose. Here's the empirical loop that compounds measurement into a flywheel, and why without it every change is a guess.
- 05 · Eval-Driven Development TDD said: write the test first, then the code that passes it. EDD says the same thing one level up — write the eval first, then build the AI system that scores against it. Here's the discipline, and why Meta's just-in-time testing is the same idea arriving from the other direction.
Agentic Foundations4 posts
The on-ramp — read this first.
- 01 · Terms I Need to Know The whole catalog leans on a handful of words and assumes you already have them. This is the on-ramp's on-ramp — an opinionated field guide to the vocabulary you need before the rest of the series lands.
- 02 · What Agentic Development Actually Is Everyone says 'agentic' and nobody defines it. Here's the line that actually separates it from chat, copilot, and autocomplete — the agent takes multi-step action, in a loop, against a goal you set.
- 03 · Context Is King Every post in this series leans on one word and never defines it: context. It's not the setting for the work — it's the medium. Here's what a context window actually is, and the four moves that separate engineers who get structured output from ones who get slop.
- 05 · Don't Just Give Your LLM Tools — Give It Systems A tool bolts one capability onto one agent for one session. A system teaches every agent you'll ever connect. Here's the difference, and why the vault — not the tool list — is the real leverage.
Sustainable AI3 posts
The honest tradeoffs.
- 01 · The Dark Side of Agentic Development: When Your AI Army Burns You Out AI agents promise infinite productivity. But when you are the orchestrator of 50+ agents across multiple runtimes, the cognitive load can be crushing. Here is what nobody talks about.
- 02 · The Hidden Cost of Building the Future: Executive Burnout in the Age of Agentic AI Why the leaders building autonomous AI systems are uniquely vulnerable to burnout—and what the data reveals about protecting yourself in the agentic era.
- 03 · What happens when the party's over. Token prices have been falling for two years. They will not fall forever. Here is how I am architecting agentic systems to survive the day one provider stops being economical.
PCMS2 posts
Personal context management, systematized.
- 01 · Your AI Doesn't Know You: The Case for a Personal Context Management System LLMs are powerful but context-blind by default. A PCMS — your notes, daily logs, and projects, structured so an AI can read them — is the real leverage in an AI-first world.
- 02 · The PCMS Manifesto: Building a Personal Context Management System Why Personal Knowledge Management Systems fall short in the age of AI agents — and how a Personal Context Management System built on Obsidian, Qdrant, and Supabase changes everything.