Thesis — May 2026
Agents, cognitive debt, and the future of governed work.
Everyone is worried about the wrong thing.
The conversation about AI agents has mostly been about autonomy: how much should agents do on their own? When should humans stay in the loop? What happens when agents make mistakes?
These are real questions. But they are not the central question.
The central question is not whether to delegate. It is how delegation accumulates debt.
Humans have been delegating thought for as long as we have had tools. Writing is delegated memory. Spreadsheets are delegated arithmetic. Search engines are delegated recall. Email is delegated presence.
Each wave of delegation created new dependencies, new failure modes, and new forms of expertise. The people who understood the failure modes became valuable. The people who only used the tool became fragile.
Agents are the next wave. They are not special because they are autonomous. They are significant because the surface area of delegation is suddenly enormous.
When you delegate a task to an agent, you get time back. That is the obvious gain. The less obvious cost is that you also give up the friction that kept your understanding current.
Friction is not always waste. Doing a task yourself, even inefficiently, keeps you calibrated. You notice when the world has changed. You build intuitions that generalize. You catch the edge cases before they become incidents.
Delegation without governance creates cognitive debt. The agent does the task. You lose the signal. Over time, the human's model of the world drifts from reality, while the agent's outputs become the inputs for more agents.
This is not a hypothetical. It is already happening in sales, marketing, and support organizations that have moved to AI-generated content and outreach. The volume goes up. The calibration goes down. When the model changes, the humans do not notice until it is too late.
Go-to-market is where I see this most clearly because it is where the pressure to delegate is highest and the feedback loops are longest.
A sales team that delegates outreach to an agent gets more volume. But they lose the signal about what resonates. They lose the conversations that would have told them the market had shifted. They lose the friction that makes salespeople good.
A content team that delegates writing to an agent ships more. But they lose the editorial judgment that made the content work. They lose the debate about what matters. They lose the craft.
The problem is not delegation. The problem is undiscerned delegation. Delegation without a system for preserving what matters.
The answer is not less delegation. The answer is delegation with better infrastructure.
What does that infrastructure look like? It looks like the systems I am building at Scale Intelligence:
Memory systems that preserve context across agent interactions, so the delegation does not also erase the organizational knowledge that makes delegation safe.
Eval harnesses that measure what the agent is actually producing, not just whether it produced something. Output volume is not quality.
Retrieval systems that connect agents to the current state of the world, not just the training data they were given. Agents need fresh signal, not just fast recall.
Permission systems that let humans specify the boundaries of delegation explicitly, not just hope the agent stays in bounds.
Review loops that bring humans back into the process at the moments that matter, without requiring them to monitor everything.
Benchmarks that let organizations measure whether their delegation is working, so they can adjust before the debt becomes a crisis.
I am starting with go-to-market because the pressure is highest and the tools are worst. Sales and marketing teams are adopting agents faster than any other function, and they have almost no infrastructure for governing that adoption.
GTM is also where the feedback loops are visible. Revenue. Pipeline. Conversion. CAC. These are not soft metrics. If your delegated GTM system is not working, you will know. The question is whether you will know soon enough to fix it.
The work I am doing with GTM Arena, Deep Content, and The Agency is applied research on that question. We are building benchmarks, content systems, and agentic workflows for GTM teams that want to delegate well, not just delegate fast.
If you are building agents, the question is not just "can the agent do the task?" The question is "what governance layer does this agent need to be safe to deploy at scale?"
That means thinking about memory: what context does the agent need, and what context does it need to preserve?
It means thinking about evaluation: how will you know when the agent is wrong, and how quickly will you find out?
It means thinking about permissions: who can the agent act on behalf of, and for what?
It means thinking about review: at what points in the workflow does a human need to be in the loop, and what do they need to see when they are?
These are not nice-to-haves. They are the difference between delegation that compounds value and delegation that compounds debt.
The problem is not delegation. The problem is the absence of the operating layer that makes delegation trustworthy.
That is what I am building.
Thierry Damiba — Founder, Scale Intelligence. Developer Evangelist, Arcade. San Francisco.