Mohamed Douare Logo

← Blog

Series · 12 articles, 2 published

Software that acts

This series explains, step by step, what software is becoming, and what I built to check it.

Each article takes a few minutes to read, explains every term the first time it appears with an everyday picture, and has a figure you can play with: a step-by-step walk-through, a chart or a small simulator. The numbers come from my own projects, measured between 8 and 17 November 2026.

The path has five stages: understand, stay free of any one provider, keep agents in check, open up and operate, then what comes next.

Start with part 1 →

1. Understand

  1. Software that acts: what software is becoming

    Today, you open an application and you click. Tomorrow, you express an intent and agents do the work. Here is what changes in simple words.

  2. An agent is a model plus a harness

    The model proposes. The harness is the program around it: it decides what is allowed, runs it and keeps a trace. Let us open the hood, with a real run.

2. Stay free of any one provider

  1. What lasts is ours; models plug in

    How to depend on no model provider: a gateway, roles instead of model names, and a real measurement of the same task along five paths.

  2. A day without Claude

    Changing provider is proven by doing it. Three real tasks, coded by another agent and another model, reviewed by Claude, for a total cost of $0.34.

3. Keep agents in check

  1. Guarantees live outside the agent

    A rule written in an agent's instructions can be ignored, whereas a tool that does not exist cannot be called. How I isolate my agents, and how I check that the isolation holds.

  2. Guides, sensors and the ratchet

    Guides are what the agent reads before acting, sensors are what checks afterwards, and the ratchet turns every error into a rule or a test. Here is how the three work together.

  3. The risk score: setting how much autonomy an agent gets

    Fixing a typo and changing a database do not call for the same control. A risk score, calculated on every change, decides whether it goes through automatically, waits for a human confirmation or is blocked.

  4. Evals: the tests of an agent

    An eval is a series of questions with the expected answers, which grades an agent the way a test grades code. Here is mine: 43 cases, 10 of them traps, once a month.

4. Open up and operate

  1. Opening a business app to AI agents, safely

    I opened the internal management platform of an international company to agents, with one rule: the agent acts as the person, and it sees what the person sees on screen, nothing more.

  2. Agent observability: see, count, replay

    While an agent is working, who watches what it does, and what it costs? Traces, four alarms, and a replay that reproduces a session without calling the model again.

  3. Bringing AI infrastructure under Terraform, without breaking anything

    To open an application to agents, you have to create accounts, groups, secrets and protections. Here is how I described them in code, without recreating anything or destroying anything.

5. What comes next

  1. A lesson that is a simulation, not a video

    I am building a platform where every lesson is a simulation that really runs. The tutor sees what the learner has done, and no exercise is shown unless the simulation has confirmed the answer.

A question, or a subject you would like me to cover? Send me an email.