In March, I wrote about what changed when I started pairing with agents: work that used to wait behind client deadlines or get lost when I moved to the next thing began moving again, technical debt became finishable, and I got farther into an idea before the rest of the day pulled me somewhere else. I still believe in every part of that shift, but I had not yet understood what it would feel like once creating the first version stopped being the hard part.
Five months later, McQueen Analytics can finish more work than it did before I began pairing with agents, while I carry more responsibility and more context in my head than I ever have.
finishing more changed the load I carry
At first, working with agents felt like getting time back because I could leave a project with enough context to keep it from starting at zero when I returned, while keeping several pieces of work moving without pretending my brain had become linear. That still feels like a real improvement in how I work.1
By late August, I had a different problem because I rarely needed another agent to produce another file; I needed to know whether the file should exist, how it fit the rest of the system, and whether I could still explain the decisions inside it. The work can become larger before I have made it better, which means the speed itself adds decisions to the part of the company I have to hold in my head.
faster software still has to fit together
Simon Willison recently wrote that cognitive capacity becomes the limit in software built with coding agents because an experienced engineer can produce much more working code without gaining the same ability to understand and maintain all of it. He connects that to conceptual integrity, or whether the parts of a system still fit together for the people responsible for it, because another room can now be added before anyone has spent enough time deciding whether it belongs.2
I know that temptation because when something used to take a week, the week itself forced me to decide whether the idea deserved that much time. Once an agent can build a first version in an hour, that question no longer arrives on its own, so I have to supply the judgment that time used to supply for me.
Linear's first report on AI use among its customers shows a version of the same problem, with agents and tools connected through the Model Context Protocol (MCP) creating just under half of new issues in Linear by August 2026. Time spent creating, triaging and commenting went up across most functions while planning time barely moved; because Linear is reporting activity among its own paid users, I am using it as a picture of what happened inside that system rather than a picture of the whole software market.3
McQueen Analytics is much smaller than the teams in that report, but the pressure feels familiar because agents can create branches, drafts, tests and possible answers faster than I can review how all of them fit together. The same speed that helps clear old debt can create new debt when each finished piece adds another decision or context trail that I am still responsible for carrying.
solo workers can inherit an executive-sized stress load
Agents have given solo workers the stress load of SVPs in companies who have to balance trusting the workforce while remaining completely responsible for its output. I can delegate more work to agents, and McQueen Analytics can finish more because of that delegation, but the company did not become less mine when the work became easier to distribute.
That pressure shows up clearly in my writing because an agent can produce a clean article in minutes with cited facts, working links and a passing score from every writing check I run, and I can still read it and know it is wrong. The piece may have no discovery, no real stakes or a first-person voice I would never use, which leaves me responsible for catching something that looked complete before it reached me.
I built Voice of McQueen from my own writing and corrections because I want the tools to step into me, producing something I could have written and can still stand behind, instead of returning generic agent copy that only sounds human after it has been polished. I call that augmented voice because a clean score cannot rescue words I would never use.
Software creates the same responsibility because passing tests and a clean review can tell me whether a change works as described without deciding why it exists, how it changes the product or whether I will still understand it when something fails later.
every piece of work has to carry its context
I have spent much of this year building systems around the agents because a conversation does not hold enough by itself, so each piece of work needs a reason, an owner, a source, a next decision and a record of what changed. The agent needs enough context to act while I need enough context to understand what comes back and decide whether I am willing to carry it into the company.
This adds more structure around tools that were supposed to save me time, and that tradeoff is real because skipping the structure saves time at the beginning while forcing me to reconstruct the work later. Repetitive, low-risk work can move quickly, while a public claim, a company system or a decision that is hard to reverse needs enough context for me to understand the cost of getting it wrong.
delegation never moved the responsibility
In a recent conversation about AI engineering, Addy Osmani called it cognitive surrender when people lose their understanding of the problem and the decisions being made; his point was that I do not need to read every token an agent produces, but I still need to understand the important decisions and answer for the part of the system I own.4
If a post says something I do not believe, a tool reaches farther than it should or a system becomes impossible to explain, I cannot point to the agent because McQueen Analytics is still my company. Output can arrive looking complete with headings, code, tests and confidence, and I still have to explain what evidence supports it, why it entered the company and what I will do if it fails.
working with agents changed my job with it
By August, working with agents had increased how much McQueen Analytics finished and how many pieces of work I kept moving at the same time, which is exactly why I still want the speed and have built much of the company around using it well. Every finished branch, draft or system also adds something I need to understand, defend or repair, so the mental load grows even when the production time falls.
I am responsible for building systems that let me trust the work without pretending responsibility moved somewhere else, because the real cost of finishing more is that there is now much more for me to understand and carry.
Other reads on AI output, context and responsibility
- Simon Willison's essay on conceptual integrity and lines of code describes the cognitive limit that remains after agents accelerate production.
- Linear's 2026 report on how teams build measures AI adoption, issue creation, coordination and planning inside Linear's own customer base.
- The Pragmatic Engineer's conversation with Addy Osmani covers cognitive surrender, mutual learning and accountability in AI-assisted engineering.
Source notes
- Carl McQueen's March 16 post “systems outcomes” and March 19 post “story and workflow reality: adhd, parallel tracks, and finishing more” described continuity, finishable technical debt, parallel work and the areas where Carl's own context and judgment remained necessary.
- Simon Willison's August 19 essay “Conceptual integrity and counting lines of code” argues that agentic coding can increase production while leaving cognitive capacity as a limit. He uses conceptual integrity to describe software whose parts fit together and notes that cheaper implementation removes some of the time pressure that once discouraged weak additions.
- Linear's 2026 report “How teams build: AI usage patterns in software teams” says its evidence comes from activity inside its own customer base and does not capture AI use outside Linear. Its report says agents and MCP clients were creating just under half of new issues by August 2026; time spent creating, triaging and commenting increased across most measured functions from June 2025 to June 2026, while recorded planning time changed little.
- The Pragmatic Engineer's August 19 conversation “From Chrome DevTools to AI Engineering, with Addy Osmani” describes cognitive surrender as erosion of comprehension and memory around the problem. It also records Osmani's view that people should understand the major decisions made with an LLM and that accountability can remain with an engineer who owns and reviews a system even when that person did not write every line.