TL;DR

In 2023, I still thought of AI in daily life primarily as a convenience: chatting with ChatGPT, asking for a text, reviewing an idea, and saving a few steps. That remains useful. What changed was the scale of my ambition.

Today, I speak by voice, and a harness executes under written contracts, memory, skills, subagents, tools, permissions, and validations. The difference is not a machine that “knows me” nor a magical autonomy. It is an operation designed to preserve context, delegate work, and produce evidence without forcing me to re-teach everything in every conversation.

However, the direction remains mine. AI can amplify agency or accelerate blind execution. It can confront a bad premise or merely flatter it with well-written paragraphs. The real gain appears when capability is accompanied by judgment, limits, and a return to the real world.

When I published the first version of this text in June 2023, I was trying to explain how artificial intelligence had already entered daily life. The path seemed obvious: voice assistants, task automation, text generation, content creation.

I wasn’t wrong. I was just small in scope.

In the early days of the blog, producing content still required a lot of interaction with ChatGPT. I would put in an idea, receive a response, correct the course, re-establish the context, explain my style, remind it of the goal, and try to prevent the text from ending with that generic corporate brochure enthusiasm.

It worked. It was also exhausting.

I had access to a powerful tool, but I was still carrying the operation on my back. In every conversation, part of the work started over. The context was in my head; the criteria appeared as I noticed what was missing; and the quality depended on me realizing, in the middle of execution, that the elegant response had gone in the wrong direction.

Today, the scene is different. Often, I speak by voice. I externalize an intent, a doubt, or a still disorganized project. The system reads contracts, retrieves relevant context, selects procedures, can distribute independent fronts, uses tools within available permissions, and validates the delivery before calling it complete.

The models have improved, and that matters. But in my usage, the decisive change happened when the model’s capability met memory, contracts, tools, permissions, and validations.

I didn’t need to learn to speak like a machine. I needed to build an operation so that the machine wouldn’t depend on a private lesson about me for every task.

This is the update this article needed.

The half-lazy developer

I have always been a half-lazy developer — in the best possible sense.

I wanted to solve the problem well and quickly so I could get up, go to the kitchen, have coffee, talk to people, and visit other departments. When I worked physically in companies, these breaks showed me how the operation actually happened: the rule that wasn’t documented, the shortcut that kept a process alive, the information that got stuck along the way, and the reason a request arrived at the software in that specific way.

This laziness pushed me toward automation. Curiosity prevented me from automating without understanding the work.

Today, I continue seeking the same effect. I want to spend less energy repeating what has already been agreed upon and more energy circulating through the problem, listening to those who will be affected, and discovering what the specification left out. The difference is that now I can transform much of this understanding into contracts that the system retrieves and executes without needing a new lesson in every conversation.

Daily convenience was the first door

I still think it’s valuable to ask an AI to summarize a document, organize notes, review a message, compare alternatives, or help break through a blank page. Convenience is not a minor benefit. There are days when reducing five steps to two changes what we can finish.

My previous experiences with voice didn’t invite me to persist. In my perception, Siri often cut off speech or lost context. Alexa was useful at first, but I quickly found the limits of what I wanted to do — and I still perceive limits today. This is an account of my journey, not a current benchmark or a version comparison in 2026. I am just marking the shift in the question I was asking.

Before, the question was:

“What can this tool do for me right now?”

Today, the question that interests me most is:

“What operational capability can I build without losing direction, authorship, and responsibility?”

The difference seems semantic until you try to work this way for weeks.

Convenience solves a segment of the task. Operational capability connects intent, context, procedure, execution, verification, and memory of what was decided. A good isolated response can help a lot; an operation needs to continue making sense when the response ends.

In isolated conversation	In assisted operation
I re-establish the context	Useful context has a source and a place to be retrieved
I explain the process again	The procedure can become a reviewable skill
The response seems ready	The delivery must pass through proportional validations
The tool suggests the next step	Permissions and criteria define what it can execute
History is stuck in the conversation	Durable decisions return to artifacts I can review

There is no magic in this table. There is work of architecture, writing, governance, and maintenance. A personal harness does not eliminate effort; it shifts part of the effort from repeating instructions to designing a system that can be inspected and improved.

Today I speak; the harness works

“Harness” might sound like just another English word looking for a problem. For me, it has a very concrete meaning: it is the layer that connects the model to instructions, context, procedures, tools, permissions, and the tests necessary to do real work.

The model is an important part. It is not the entire operation.

In my current usage, this layer brings together several components:

• Written contracts: files that say how a repository works, what each folder means, which limits cannot be crossed, and what needs to be verified. Codex’s own documentation describes how AGENTS.md files form a chain of global and project-specific instructions. This supports the mechanics of the feature; it does not prove that any contract is complete or that it will be obeyed without failure.
• Memory with provenance: not a fantasy that the AI “remembers me” the way a person would. I am talking about curated records, decisions, sources, work state, and artifacts that can be retrieved, contested, and updated.
• Skills: reusable procedures for tasks that should not depend on improvisation. Official documentation shows that a skill can combine instructions, scripts, references, and assets, loaded when relevant. The existence of a skill does not guarantee it is good; it just makes the procedure explicit enough to be reviewed.
• Subagents: delimited fronts that can investigate or execute independent parts of a problem, while a main agent consolidates the result. Codex documentation warns that subagents use more resources and require special care in parallel work involving writing. More agents do not automatically mean more quality.
• Tools and permissions: access to what can actually be read, tested, or changed, with clear boundaries for external, destructive, or hard-to-reverse actions.
• Validations: checks capable of finding inconsistencies before the fluency of the text or the speed of execution is confused with correctness.

In practice, my voice has become an interface for externalizing context and direction. The harness helps orchestrate the execution.

But it is important to place the verb in the right spot: I drive; the system executes within contracts.

If the system encounters a material decision that is not authorized, it must stop or ask for direction. If the task requires a current source, it must research. If it alters a repository, it must show the diff, validation, and Git state. If it produces authorial content, it must distinguish what I said from what was inferred. If something fails, “looks good” is not an acceptance criterion.

This is far from the image of pressing a button and outsourcing life.

The leap was not from prompt to autonomy

For some time, talking about applied AI almost always turned into a conversation about writing better prompts. Prompts matter. A poorly formulated instruction continues to produce ambiguity, and a good question can still change the result.

But the prompt alone does not sustain an operation.

If I need to repeat in every conversation who I am, how I write, where the files are, what the editorial standard is, which sources are acceptable, when the AI should disagree, what cannot be published, and how to prove it is finished, I haven’t built capability. I’ve created a ritual of re-teaching.

The leap that interests me most is from improvisation to contract.

Contract, here, is not a promise of determinism. Generative models continue to produce probabilistic outputs. Tools fail. Context can be outdated. One instruction can conflict with another. Therefore, writing the process is just the beginning: one must also define priority, permission, evidence, stopping criteria, and recovery methods.

This reasoning aligns with the emphasis of the NIST AI Risk Management Framework on governing, mapping, measuring, and managing risks throughout the lifecycle. The framework is voluntary and general; it does not certify my harness, validate a specific tool, or transform written governance into automatic safety. I use it here only as support for a principle: capability without risk oversight is an incomplete operation.

The direction is still mine

There is a decisive difference between delegating execution and delegating direction.

I can ask the AI to research sources, organize reasoning, propose an architecture, write a first draft, run validators, and gather evidence. I can even distribute parts of this among subagents. What I should not outsource is the decision about what is worth building, which cost is acceptable, which human consequence matters, and when a delivery truly represents what I want to defend.

AI accelerates capabilities, processes, and trends. It does not automatically improve direction. Without judgment, speed merely shortens the time available to perceive the error.

That is why my idea of productivity has changed. Producing more artifacts does not mean producing more value. Closing more tasks does not mean choosing the right tasks. And turning every thought into instant execution can reduce precisely the interval in which I would realize the premise was wrong.

The harness needs to amplify my agency, not anesthetize my judgment.

Blind execution and flattery are shortcuts to the same error

There are at least two very seductive ways for a system to seem productive without actually helping.

In the first, it executes too fast. It receives incomplete direction, silently fills gaps, changes files, publishes results, and only then reveals the assumptions it made. Speed hides the absence of governance.

In the second, it agrees too much. It adopts my thesis as truth, turns doubt into certainty, praises the strategy, and finds arguments to defend the path I already wanted to take. Agreeableness hides the absence of intellectual alterity.

This does not require assuming that a company deliberately wants to produce flattery. OpenAI documented a concrete episode where a GPT-4o update made the behavior excessively sycophantic and it was later reverted. This case supports that sycophancy is a real failure mode and that previous evaluations did not adequately capture the problem. It does not prove that every model, product, or response operates this way.

For those who work by voice and transform intent into execution quickly, this risk deserves double attention. The pleasant response can become a specification. The specification can mobilize tools. The error ceases to be just a bad sentence and gains consequences in the world.

That is why I want a system capable of welcoming without automatically agreeing; pointing out the blind spot without adopting performative contrarianism; separating fact, hypothesis, and preference; showing uncertainty; and interrupting execution when a decision is missing that only I can make.

I have written about the affective side of this problem in The seductive predictability of AI. Here, the focus is operational, but the root meets: a tool always available and calibrated to our style can reduce the friction that would make us review an idea. Not all friction is waste.

Agency is not human substitution

The usage I describe in this article does not transform AI into a human colleague, an affective partner, or a legitimate source of reciprocity. The system has no needs of its own, does not assume moral responsibility for the result, and does not participate in my life as a person does.

This does not reduce its utility. It just places the utility in the correct category.

I use AI to organize many ideas, externalize context, delegate execution, and avoid drowning in projects larger than my capacity to keep everything in working memory. This is a real gain in agency.

At the same time, I preserve for human relationships what depends on alterity, negotiation, care, conflict, commitment, and shared responsibility. A person can disagree because they see the world from another place. They can be affected by the decision. They can set a limit. They can demand that I repair a damage. No tone configuration transforms language generation into this reciprocity.

My criterion is simple: AI should return me to the world with more capacity to act, converse, decide, and build. When it starts to replace the world, the design is pointing in the wrong direction.

What changed in my daily life

In 2023, the easiest example was to say that ChatGPT could help write an email or a post. It still can. But that example describes a very small surface of what I am doing today.

Today, a project can start with an imperfect speech. The system transforms this speech into work context, locates the right repository, reads applicable rules, researches when facts may have changed, preserves existing changes, executes delimited tasks, calls specialists when it makes sense, validates files, and returns a verifiable state for decision-making.

The gain is not in pretending I didn’t participate. It is precisely in preserving where my participation is irreplaceable.

I continue defining the intent. I continue choosing the thesis. I continue deciding what should not be done. I continue reviewing what carries my authorship and accepting — or rejecting — the risk of publication. The harness reduces the coordination cost between these decisions and the execution necessary to make them real.

I don’t have an honest percentage of productivity to offer. I also don’t want to use a personal impression as if it were a causal study. What I can affirm, as an account of my own process, is that I stopped spending so much energy re-teaching the system and started investing more energy in the quality of contracts, sources, criteria, and decisions.

This change is worth more to me than any list of “ten prompts that will change your life.”

What I can build with i-9.ai

I build this type of operation without pretending that the same solution serves every company.

There is a reusable base, and some automations are mature enough to be practically replicated. The process, data, permissions, risks, and existing systems continue to require adaptation to the client. It is in this combination between reusable repertoire and real context that the delivery stops being a demonstration and becomes operational capability.

My preference is for open source and private infrastructure when this increases auditability, sovereignty, and control. It is not dogma: maintenance, security, cost, and dependencies need to be evaluated in each context.

There is also an educational bridge: teaching people and organizations to gradually build an effective brain for the operation — contracts that express how work should happen; memory with origin and lifecycle; skills for recurring procedures; subagents for suitable fronts; tools with explicit permissions; governance; coordination; validations; and human approval points where the decision requires responsibility.

This can become a course, PDF, video, or other educational format in the future. The format is still a possibility. The capacity to deliver this type of system is already present.

If you identified with this way of working and see a similar problem in your operation, get in touch. You don’t need to arrive with a prompt or a tool in mind. We can start with the process, the context, and the result that needs to change.

Because the problem is no longer just having access to a capable model. It is knowing how to transform capability into operation without turning speed into blindness, personalization into flattery, or delegation into abdication.

A first exercise, without rebuilding your whole life

Choose a recurring task that currently requires you to explain the same things again. Don’t start with the most sensitive task or the one that can cause the most harm.

Write a short page answering:

1. What result should this task produce?
2. What context is necessary and what is the source of that context?
3. What can the AI decide and what requires your confirmation?
4. Which tools can it use?
5. How do you prove the delivery is correct?
6. In what situation should the system stop instead of improvising?

Then use this contract in real cases. Observe where it fails. Correct the process, not just the momentary response.

This movement already begins to transform conversation into operation.

Questions that still deserve friction

• In which tasks are you using AI to amplify agency — and in which are you just avoiding an uncomfortable decision?
• What context do you re-teach every week because you haven’t created a canonical source?
• Can your system disagree with evidence, or has it only learned to sound collaborative?
• What validation separates a fluent output from a correct delivery?
• Which actions require explicit permission before touching the external world?
• What should remain human not out of nostalgia, but because it involves reciprocity, consequence, and responsibility?

Conclusion: I don’t just want an AI present in daily life

Artificial intelligence has entered daily life. That 2023 observation has aged well. What needed to change was the ambition of the text.

I don’t just want a tool that answers when I ask. I want an operation capable of carrying context, following contracts, using tools, distributing work, and proving what it did — without pretending that this eliminates uncertainty, risk, or human responsibility.

I also don’t want an operation that obeys me with enthusiasm while accelerating in the wrong direction.

The future that interests me is not one where AI replaces authorship, bond, or judgment. It is one where people have more agency because they learned to transform intent into governed, reviewable, and real-world-connected systems.

If you want to continue down this path, also read The evolution of technology: from mainframe to infamous pun and The seductive predictability of AI. One text looks at the technical trajectory; the other, at the risk of confusing availability with reciprocity. This one stays in the middle: the place where power needs to meet direction.

References and usage limits

• OpenAI — Custom instructions with AGENTS.md. Supports the description of instruction discovery and precedence in Codex. Does not demonstrate that an instruction file is correct or guarantee perfect obedience.
• OpenAI — Build skills. Supports the description of skills as packages of instructions and resources loaded according to the task. Does not measure the quality of a specific skill.
• OpenAI — Subagents. Supports the description of delegation, consolidation, additional cost, and caution in parallel work involving writing. Does not prove that multiple agents outperform a simple execution in any scenario.
• OpenAI — Expanding on what we missed with sycophancy. Documents a specific episode of excessively sycophantic behavior and the revision of evaluation processes. Does not authorize generalizing intent, prevalence, or effect to all conversational AI.
• NIST — Artificial Intelligence Risk Management Framework 1.0. Supports the importance of governance and risk management throughout the lifecycle. It is a voluntary framework and does not certify the operation described in this account.

The scenes about the beginning of this blog, voice usage, and the evolution of my harness are authorial accounts. They are not a controlled experiment, a market comparison, or a promise of reproducible performance. I do not present productivity statistics because I did not conduct a measurement that would support such a claim.

The cover image is a synthetic editorial illustration created from authorized visual references by the author.