Your Job Title is a Liability

## Editorial

Last month I wrote about [Human Agentcy](https://www.thatwastheweek.com/p/human-agentcy). The spelling mattered. I was trying to describe something more specific than "agents" and less passive than "automation." Agentcy is human agency extended through software that can act. The point was not that AI replaces human agency. It was almost the opposite.

Extension, not replacement. Delegation, not autonomy.

That now feels like the center of this week.

The first phase of generative AI was about tools. A chatbot could write, summarize, code, search, explain, or draw. The second phase is about agents. Agents do not just answer. They enter workflows, use tools, make calls, run loops, invoke other agents, and return later with results. They turn software from something we operate into something that operates on our behalf.

That sounds like the human role should get smaller. I think it gets bigger.

The skill requirement goes up, not down.

Google’s Sundar Pichai made the interface change explicit. The old internet was built around browsers, search boxes, tabs, forms, and pages. The agentic internet is built around delegation. Instead of visiting a website, finding the right page, filling out 18 fields, uploading a document, and checking back later, an agent does the chore.

The line that stuck out was:

“Once you get the taste of the superpower that comes with agents... you realize they aren’t just tools, they are agency.”

That is the right word. Agency. But it is not agency floating in the air. It has to be designed, constrained, audited, and improved. If an agent renews your license, patches your code, routes your customer issue, or launches a database, the important question is not whether the agent can do the task once. The important question is whether the environment around it makes the action reliable.

Databricks CEO Ali Ghodsi gave a second proof point. According to the clip circulating this week, more than 81% of new Databricks databases are now being created by autonomous agents rather than humans. That is not a forecast. It is an operating fact inside a major enterprise data platform.

His line was even sharper:

“Writing software is ten times faster now... your margin is my opportunity.”

That is the old software industry hearing the new one knock on the door.

If software can be created ten times faster, barriers to entry fall. If agents are the primary users of a platform, product design changes. If workflows can be assembled and reassembled by small teams, coordination becomes the bottleneck. The advantage moves away from simply owning a large application and toward owning the loop, the data, the judgment, and the operating system around the agent.

That is why Dan Farrelly’s “Agent Loop Architecture” is more important than it may first appear. It is not merely a developer note about infrastructure. It is a map of the new work.

Farrelly says the agent stack has three layers: loop, skill, orchestrator. A loop is a repeated timed job, plus a decision-maker. A ‘skill’ is a durable workflow. The orchestrator is the engine that checkpoints steps, retries failures, stores run history, enforces concurrency, and lets new functions deploy without breaking work already in flight.

His cleanest summary is this:

“The loop is plumbing. The asset is the skill it calls.”

And the operational requirement is just as plain:

“These things have to survive a restart.”

This all sounds technical, but it is a job definition. A new job. This isn’t an engineer coding. This is a human designing workflows carried out by human-designed agents. If you can walk in to any company and learn, and then automate, its workflows, you have the skills for the new job. It involves learning, listening, designing, articulating and delivering actual outcomes. And things are moving so fast, you will redo the whole thing in six months tops.

In the old software world, skill meant knowing the app, the workflow, the data, and the business rule. In the agent world, skill means knowing what should be delegated, designing the loop, choosing the right model and tool mix, setting constraints, auditing traces, catching errors, and improving the skill library over time.

The valuable worker is no longer the software engineer, or the operator of software. It has transformed into the designer and governor of agent environments.

That creates new jobs, or at least new definitions of existing jobs. Agent operator. Loop designer. Skill builder. Agent ops. Eval designer. Context architect. Model router. Human governor. These sound like awkward titles today because we are at the beginning of the transition. So did “social media manager” once. So did “cloud architect.” So did “data engineer.”

The deeper point is that the best human work lives outside of the code base and lives in strategy and execution. The person who once completed a workflow now designs the loop that completes it. The person who once checked a report now defines the evaluation that tells whether the report is good. The person who once remembered how the company works now encodes that memory into a reusable skill that survives a model swap, a process restart, or a change in vendor.

Satya Nadella’s phrase for this is “token capital.” The phrase is useful because it points to a new form of institutional wealth. A company has human capital, the knowledge and judgment accumulated by people. It also now has token capital, the workflows, traces, prompts, evals, skills, permissions, and agent loops that capture how work gets done.

The moat is not the model. Models will change. Prices will change. Capabilities will change. The moat is the learning, and permanent updating system, around the model.

Anthropic, Databricks, Z.ai, Google, Mutiny, and the open-source model companies all belong in the conversation. They are not separate AI stories. They are pieces of the same migration.

Anthropic is trying to control the safety and use boundary around frontier models. Z.ai is pushing long-horizon open models into coding-agent territory. Google is turning agents into an interface layer. Databricks is seeing agents become real users. Mutiny is refounding a SaaS company around an agent rather than a UI. Farrelly is describing the orchestration layer required to make all of this production-grade.

The question underneath all of it is simple:

Who owns the overview, and the loop?

Does the model company own it? Does the cloud own it? Does the app own it? Does the enterprise own it? Does the individual own it?

That question may matter more than who has the best chatbot this month. And the answer is - one or more humans own it. With new job titles.

There is another human lesson here. Ihtesham Ali’s post about Sam Altman’s Stanford remarks captured an uncomfortable point. Altman said the experts who were most certain scaling would not work were often the ones most unable to update when the data changed. The problem was not intelligence. It was identity.

The line I would keep is:

> “The moment a belief becomes who you are, it stops being something you can update.”

That applies directly to the agent era.

If your identity is “I am the person who does this task,” an agent feels like a threat. If your identity is “I am the person who knows how this outcome should be achieved,” an agent becomes leverage. If your identity is tied to the current workflow, you defend the workflow. If your identity is tied to judgment, you redesign it.

That is why the new skill is not just using AI. It is staying updateable, or being open to continuous learning.

The AI era will make low-skill use easier. Anyone will be able to ask for a summary, a spreadsheet, a landing page, or a piece of code. That is real. But high-skill use will become more valuable, not less. Someone has to know whether the summary is faithful, whether the spreadsheet matters, whether the landing page sells, whether the code is safe, whether the loop should run, whether the permission should be granted, and whether the system is improving or merely doing more things faster.

This is where Paul Graham’s essay this week aligns with the agent discussion. Graham says startup wealth can be earned when founders make something users like enough to tell their friends about it. He ends with the claim that for startups, “the key is not exploitation but empathy.” Knowing what is wanted, and needed.

That remains true in an agent world. The best loops will not be the ones that simply replace human action. They will be the ones built by people who understand what other people need, where judgment belongs, where automation helps, and where it must stop.

Human agency does not disappear when agents arrive. It becomes more architectural.

The job is no longer just to use software. The job is to define what software should be allowed to do. And then to build the loop that lets it do it well.

Your job title is a liability. Your skills are growing. Redefining yourself is the most important small change you can do.

---