Alex Herrero

Becoming the assistant

Fri, 24 Jul 2026 00:00:00 +0000

Agent M started as something I installed in a repo, and it kept growing, so it’s moving to the host level and … slowly … becoming the assistant I had in mind when I started.

The agent couldn't click Allow

Mon, 13 Jul 2026 00:00:00 +0000

The other day I started to notice a pattern. My coding sessions kept getting stuck. No obvious errors, just stuck. I’d leave my coding session going when I went out for the day, I’d turn on remote mode, and I’d see that it would get stuck and misbehave in some way or not be able to do what I was asking. It was really weird. After a few days trying various fixes, like splitting the jobs into smaller chunks, I put two and two together: whenever this would happen, I’d come home and find my Mac prompting me to allow Claude access to a folder. And then it clicked… somehow Claude needed access and then wasn’t able to do what it had planned, and my session would start misbehaving. What was happening? The first thing any session does as part of Agent M is read its memory vault (a local storage directory). That folder needed permission. When I’m there, I click Allow without thinking about it. But a lot of the time I wasn’t there. Granting broader folder access made the prompts go away, not a perfect solution, but better. So today I learned that when something gets stuck for days, it’s worth asking what it was waiting for.

Google Drive sync has no undo

Fri, 26 Jun 2026 00:00:00 +0000

When I first built Agent M, some research led me to host my Agent Memory Vault in a Google-Drive-synced folder. It was a simple and effective way to make the vault accessible across devices — and, with some work, even reachable from Claude and Gemini on the web. Over time, though, this decision has come back to bite me a few times. Drive has no version history, just a trash bin. I recently had an overwrite I couldn’t walk back, and the odd sync-conflict file left behind after two machines wrote at once. And because I also use Obsidian to work in the vault, I had to do some gymnastics to keep Drive from syncing its index. So today I learned that “it syncs” is not “it’s backed up” — and I made plans to move the memory under git, to finally get history, diffs, and an undo.

Running Agent M all day

Thu, 18 Jun 2026 00:00:00 +0000

Running an agent all day used to mean hitting the usage ceiling by mid-afternoon. Here’s how I cut a heavy day ~90% and made Agent M cheap enough to just leave running.

The multi-surface wall

Sat, 06 Jun 2026 00:00:00 +0000

The dream for Agent M is one agent everywhere — same personality, same memory, same workflows, whatever tool I happen to be in. In practice I get most of the way there: it runs the same across Claude Code and Antigravity, and can read its memory from a couple of other surfaces. But it’s not there yet. Chat-only assistants can’t touch live files, so they’re out for actual work. And not every tool lets a setup like mine install its own hooks or scheduling — on some, my guardrails can watch but can’t act. So today I learned that “works everywhere” has an asterisk right now: the agent travels, but the surfaces underneath are uneven, and the portable, do-anything agent is getting closer, but probably still a harness generation away.

Building Agent M

Sun, 24 May 2026 00:00:00 +0000

I set out to rebuild a trading engine. What I built instead was Agent M — an agent with a durable memory that works across every tool I use.

A 1M context window isn't free

Fri, 01 May 2026 00:00:00 +0000

So agent harnesses like Claude and Antigravity recently introduced 1M-token context windows, which was great, until I realized that I was torching my token budget. Why? It turned out that the whole context window for the conversation is re-sent on every turn, so a near-full window burns through your usage limit fast. There’s a fix though — Claude has a setting to cap the context window back to the standard 200k. That significantly improved my token consumption. So today I learned that more context isn’t free — sometimes you want less of it.

Rebuild, not refactor

Sat, 18 Apr 2026 00:00:00 +0000

Sherwood came together fast, but it had drifted too much and needed a re-write. Re-writing it was where I realized that I needed to rebuild something else.

Don't get too far ahead of the agent

Wed, 25 Feb 2026 00:00:00 +0000

When I started Sherwood I designed the whole thing up front — including the front-end, before I’d built any of the backend. It felt productive. I expected that the agent would just be able to code until it was done, but by the time we got to the front-end, the backend had already drifted from the plan, and I had a pile of drift to reconcile that I should have caught earlier. You hear a lot of the term ‘Human in the Loop’ posed as a solution for this problem, and it’s true that it helps, but this also slows development, and I don’t want to be a bottleneck. So today I learned that designing too far ahead is its own kind of drift. A human in the loop catches it, but agents write faster than one person can review, so keeping pace will eventually have to be automated.

Building Sherwood

Wed, 18 Feb 2026 00:00:00 +0000

I started Sherwood to learn how markets work, and it came together fast — fast enough that the drift started getting ahead of me.

When my Claude allowance got cut

Wed, 04 Feb 2026 00:00:00 +0000

For a couple of months, I worked almost entirely in Antigravity. I could see all the code as it was being edited in real time alongside plans that I could comment on. The workflow felt like real collaboration. One of the things I liked most was that it supported both Gemini and Claude. I could switch mid-task depending on what I was doing — reach for one model where it was strong, then the other — without leaving my setup. It was a nice way to work. Then the Claude allowance in Antigravity got brought down to a level I couldn’t really use, even for ultra subscribers, and the easy switching was gone. That forced me to rethink my whole harness and how I worked. It turned out for the better: I leaned into Claude, and for what I was doing it held up better than I’d given it credit for. I still use Antigravity for the things Gemini is good at, but now I follow a ‘best of breed’ approach. So today I learned that the limit I didn’t ask for pushed me to a better setup than I’d have chosen on my own.

Green squares

Mon, 19 Jan 2026 00:00:00 +0000

For a few years my GitHub graph was mostly gray — after all, I spent most of my time managing a growing team and stack, not building. Late 2025, the tools started getting really useful, and it was time to try some new things.

Before it had a name

Sun, 15 Jun 2025 00:00:00 +0000

How a one-week work demo became a real cross-platform agent — and convinced me agents that act for you were close.

Intro -> Next Level

Tue, 10 Dec 2024 00:00:00 +0000

Six months ago, AI coding let us build an on-device agent — in just a week — that would have taken months. I knew then that software was moving to the next level — today we call it agentic coding. Here’s where I decided to write about what that’s actually like.

About

Mon, 01 Jan 0001 00:00:00 +0000

I’m Alex Herrero — a security-minded engineering leader. By day I lead engineering for AI-driven threat detection at Google Cloud. Off the clock I build the tools this site is about: Agent M, a harness that gives coding agents durable memory and a phase-gated workflow, and Crickets, the safety and verification primitives that ride on top — kill-switches, fresh-context evaluators, guardrails.

This blog is where I write about the experiments behind those tools: what I’m trying to make AI agents reliable, safe, and genuinely useful, and what I learn running them against real work.

Now

Mon, 01 Jan 0001 00:00:00 +0000

What I’m focused on at the moment (a /now page).

Building Agent M and Crickets in the open — and writing up the experiments here.
Sharpening the intersection I care about: making AI agents reliable, safe, and useful, from a security background.
Reading (and occasionally posting in) the AI-safety corners of the web.

Last updated June 2026.