Antony Evans Blog

How to build a portable memory for AI agents

Fri, 09 Oct 2026 00:00:00 GMT

Once models are cheap and interchangeable, the valuable part of an AI setup is your context. A one-minute explainer on keeping that memory in files you control, outside any one vendor.

AI cost per task: why AI feels pricier as it gets cheaper

Mon, 28 Sep 2026 00:00:00 GMT

Frontier AI runs cost more every year, but at fixed quality AI cost per task is falling about 13x a year. Why both are true, and the rate I'd plan agents on.

Zuckerberg being the savior of humanity wasn’t on my bingo card for 2026

Fri, 25 Sep 2026 00:00:00 GMT

Zuckerberg's AI glasses vision, the UN's AI standards debate, and who gets to decide the future of AI.

I got bitter lessoned

Thu, 24 Sep 2026 00:00:00 GMT

Sol-6 and Opus 5.5 broke my dev process. I'm ripping out the heavy software factory I spent three months perfecting and going agile: continuous testing on staging, frontier models at higher reasoning, and no more independent judges.

The constraints are human now

Tue, 22 Sep 2026 00:00:00 GMT

AI capability is no longer the scarce input. What's actually holding back adoption: organizational buy-in, trusted context, action authority, energy, permission to build, and economic controls.

Agents are making a second computer surprisingly useful

Thu, 17 Sep 2026 00:00:00 GMT

Why I use Linux machines alongside my main computer: overnight work, more memory, separate access and remote connections through Tailscale.

AI makes mistakes, it's stochastic, here's how I manage those

Wed, 09 Sep 2026 00:00:00 GMT

AI is fundamentally probabilistic and will make mistakes. Seven ways I manage that, from human review to fresh-context judges to shrinking the blast radius of any single error.

Here's my current AI stack

Mon, 31 Aug 2026 00:00:00 GMT

Here's my current AI stack: open source, a dedicated memory system, multi-agent access, spec judges, evals, and as much of the process in code as possible.

Git-owned knowledge base: markdown in git is not a hack

Fri, 28 Aug 2026 00:00:00 GMT

A git-owned knowledge base is markdown in a repo you control. A YC launch listed that as a hack. You want a record you can diff, refuse, and take with you.

33 workflows, five layers — what a real harness looks like from the inside

Tue, 11 Aug 2026 00:00:00 GMT

The piece I wrote on harnesses described five component layers that every production agentic system needs: orchestration and planning, an execution environment, validation and quality control, context and memory management, and logic and interaction.

Launching Shop CLI

Tue, 11 Aug 2026 00:00:00 GMT

Shop CLI: A Tool Built for AI Agents, Not Humans

How AI Agents Reshape Business Relationships

Tue, 11 Aug 2026 00:00:00 GMT

Not all business relationships do the same job. Research shows which ones survive the agentic economy, and which were always just expensive due diligence.

Meta-prompting — the thinking skill that makes AI useful

Tue, 11 Aug 2026 00:00:00 GMT

The skill that actually makes AI useful: meta-prompting

When to fix AI's work by hand vs fix the system that produced it

Tue, 11 Aug 2026 00:00:00 GMT

The short version: When an AI agent gets a task 80% right, you can patch the last 20% by hand and ship today, or fix the instruction behind it and never touch that task again. The patch is faster now. The fix compounds. Most people, me included, pick the patch too often, because the economics that should govern the choice have changed and our instincts haven't caught up.

Fable is back — what using an expensive, deeply capable model teaches about discriminating AI use

Tue, 11 Aug 2026 00:00:00 GMT

Fable is back. And it's changing how I think about thinking.

Why Your AI System Is Never Done

Tue, 11 Aug 2026 00:00:00 GMT

Antony Evans argues that building an AI system is farming, not hunting: there is no finished, shipped state, only a system you keep tending as models change.

AI Agent Burnout: Running a Team That Never Clocks Out

Tue, 11 Aug 2026 00:00:00 GMT

Antony Evans on why AI agents that work all night push founder-operators toward the same always-on habits, and why he has not solved it yet.

The honesty test for AI-assisted writing

Tue, 11 Aug 2026 00:00:00 GMT

The honesty test for AI-assisted writing isn't who typed the words, it's whether the idea is genuinely yours: a first-person account of one writer's pipeline.

I asked my AI to audit its own product's code for what was slow. It blamed my safety rules, not the code.

Mon, 20 Jul 2026 00:00:00 GMT

I asked my AI to audit its own product's code for what was making development slow and expensive. I expected it to blame the code. It blamed me.

I published the first customer-facing release of my company mind, and the hard part was making sovereignty boring and true

Sat, 18 Jul 2026 00:00:00 GMT

This week I gave my company mind away as something anyone can actually install, and the hardest part was making it boring.

The vendor should not be able to reach into your data — and my architecture makes that true

Fri, 17 Jul 2026 00:00:00 GMT

I spent today deciding what my own product is allowed to reach. The answer I kept landing on was: nothing of yours.

Someone else graded the code my AI wrote

Thu, 16 Jul 2026 00:00:00 GMT

I had someone from outside look at code my AI wrote this week. Not the output, the actual work. Every file, every test, every corner where a tired human would have cut one.

I let my company mind change itself, but only under governance

Wed, 15 Jul 2026 00:00:00 GMT

For weeks my company mind could read itself but not change itself. Today I let it write back, and the hard part wasn't the writing. It was deciding what it's allowed to change without me.

The night my own system refused a direct order

Tue, 14 Jul 2026 00:00:00 GMT

Tonight my own system refused to do what I told it to.

The best investment right now might be the founder you'll never see pitch

Fri, 10 Jul 2026 00:00:00 GMT

Last updated: 2026-07-05

AI-redpilled: the test that separates talk from what's built

Fri, 10 Jul 2026 00:00:00 GMT

AI-redpilled gets claimed constantly and checked rarely. Here's the screen-share test a Laurel CPO uses to tell real AI use from talk, for people and teams.

I gave an AI the keys to merge my own code

Sat, 04 Jul 2026 00:00:00 GMT

Today I did something that would have sounded reckless to me a year ago. I gave an AI permission to merge my own code into my repositories without asking me first.

What my second child taught me about freedom

Sun, 21 Jun 2026 00:00:00 GMT

I used to think kids were the end of my freedom. Two of them in, I've found almost the exact opposite.

Declinism is a choice, and a self-fulfilling one. Abundance mentality as the rational, generative bet.

Fri, 05 Jun 2026 00:00:00 GMT

Stop telling me the world is ending. You don't know that. Neither do I.

500 agents at Edge Esmeralda — the missing layer is discovery and mediation

Thu, 04 Jun 2026 00:00:00 GMT