The tool I use now is not the most capable one I tried. It is the one that fits how I actually work.

I do not enjoy switching tools

Every tool change costs time. New setup, new habits, new ways things break. So I only move when a limitation gets in my way during real work.

I lead marketing and digital at Cove. Before that, Traveloka, WPP, and GudangAda. I run a team, cover several markets, and keep a few side projects alive in the gaps of the week. Time is what I never have enough of. A tool earns a switch only when staying put costs more time than moving.

That is the lens for everything below. Not which tool is best. Which tool costs me the least over months.

The journey, briefly

ChatGPT came first. Good for thinking out loud. Weak at doing work in my actual files.

Then Cursor, and later Antigravity. Both strong for code. But I spent more time in an editor than the work justified for someone who directs more than he builds.

Then Claude Code. This one stuck for a while. It could read a project, make changes, run commands, and hold a plan in its head. It became my default last year.

I also ran Perplexity beside it for research where I wanted sources I could check.

Then I tried OpenClaw, because it pushed harder on autonomy. It did not go well. I hit errors I could not get past, and the work did not feel smart enough yet. So I went back to Claude Code. That return told me something about my bar. Capability first, autonomy second.

Then Hermes, which is where I am now.

The moment Claude Code stopped fitting

Claude Code is capable. That was never the problem. The problem is that it is built around the session, and my work is not.

On long projects the context drifts. Details from early in the work get clipped or lost. On a big task I learned to expect a wall somewhere in the middle: a session limit, an expired login, a context reset. Each one meant re-explaining what we were doing and hoping the thread picked back up where it left off.

Lately it got harder in another way. More hallucination in the output. I got tired of reminding it of things we had already agreed on. Even after I wrote those things to memory, it kept repeating the same mistakes. A few times I made a wrong call because it had not taken in updated data. In the end I was cross-checking its work with Gemini Pro, which defeated the point of running Claude in the first place.

For a one-off job, that is a nuisance you absorb. For work that runs over weeks, it adds up. I was spending real time getting a tool back to where it already was.

What changed my criteria

I stopped optimizing for raw capability. The best single session stopped being the thing I cared about.

I wanted something different. Something that runs continuously. Something that learns about me and my projects over time. Something that handles routine work without me starting it by hand each time.

Hermes is built around that idea. It keeps memory across sessions, so context I gave it last month is still there this month. Skills accumulate, so something I teach it once stays taught. It runs scheduled and autonomous jobs, so routine checks happen on their own. It supports multiple agent profiles, so different kinds of work stay separate. It fits a set-and-forget way of working, which is the only way I keep side projects running at all.

How I actually use it

This is my setup, not a recommendation.

I have Hermes run Claude Code for the heavy lifting. Hermes sits over the top and checks the output. It catches inconsistencies and the repeated mistakes I used to catch by hand. It feels less like a tool and more like a team member that learns.

I also run a few separate bot personas. I use them to pressure-test an analysis from more than one angle before I trust it, and to tighten the final output.

Here is the twist. I said Claude Code stopped fitting. I did not actually stop using it. Today Hermes runs Claude Code and supervises it. The problem was never the tool on its own. It was running it without memory and without supervision. Under Hermes it is useful again.

The rough edges are real. Hermes is newer, so there is less written about it when something breaks, and fewer people to ask. For the work I do, that has not bitten me yet. For a large one-off coding session, Claude Code is still better, and its tooling is more mature. I did not move because Hermes wins on every axis. It does not. The gap is closing faster than I expected. Hermes on DeepSeek already handles most of my work well. I am weighing whether I still need Claude Code at all. I am not there yet.

The operator's read

Here is what I take from it. For someone who runs a team and a set of projects, the tool that wins is the one that remembers your context and lowers what you have to maintain over time. Not the one with the best demo.

Raw power is easy to sell and admire. But when your time is the real constraint, reliability and continuity are worth more than a higher ceiling. I would rather work with a tool that is a step behind on capability and never makes me start over, than a brilliant one that hits a wall every few days.

Hermes is the tool that fits how I work today. That is the only claim I am making.

How are you managing your workflow?