The more capable agents become, the more we're making them work like people.
I don't mean making them sound human. I mean giving them some of the structures humans use to make our own intelligence more reliable.
None of this borrowing is new. Neural networks themselves were inspired by biological neurons, even if the resemblance between the two is now fairly loose.
But we're increasingly borrowing from how humans think and work, not just from ideas about how the brain works.
Memory is a good example. Human memory isn't a complete transcript of everything we've experienced. We retain some things, consolidate them over time, retrieve them when they're useful and forget others. Agent memory is increasingly becoming a software process with many of the same concerns: what belongs in immediate context, what should persist, and what should be retrieved later. Research into sleep and memory consolidation has long described consolidation as part of biological memory, while experiments such as Do Language Models Need Sleep? explore analogous processes in models.
The same thing is happening with patterns that aren't obviously cognitive. Review, delegation, clarification and approval come from how humans organise work around each other's limitations.
We're not just borrowing ideas about how intelligence might work. We're borrowing what humans have learned about making intelligence useful.
Engineering as a Frontier
Software engineering is an interesting place to watch this happen. It's a form of knowledge work with unusually mature ways of dealing with the fallibility of individual judgement. We don't rely on one engineer getting everything right. We use code review, tests, CI, staging environments and approval gates.
As coding agents have become more capable, we've started giving them the same structures. They run tests and linters before presenting their work. Subagents independently investigate parts of a problem and report back to a coordinating agent. Human approval sits in front of consequential actions. Anthropic's multi-agent research system and its work on evals are concrete examples.
Matt Pocock's
Grill Me skill
is an especially simple one. Rather than handing an underspecified idea to an agent and hoping it fills in the
gaps correctly, the agent asks questions until it understands what you actually want.
The pattern isn't new. It's requirements gathering.
That fits Pocock's broader argument in Software Fundamentals Matter More Than Ever: agentic engineering still depends on the judgement and critical thinking that engineering always has.
That's what I find interesting about a lot of the progress in coding agents. We're not only making the underlying intelligence more capable. We're rediscovering the practices humans already use to make intelligent work reliable.
Beyond Engineering
Coding is a particularly visible frontier for these ideas, but I don't think that's where they stop. Increasingly, software is the medium through which agentic work gets done, not necessarily the output.
I became interested in this while prototyping an AI-assisted tool for writing career goals and development plans.
Coding gives us unusually strong ways to evaluate agent output. We have tests, type systems, linters and code review. Career development doesn't have those things. A career goal doesn't compile.
But the collaborative pattern still worked.
The AI chat and workflow inputs sat side by side. As the conversation unfolded, the agent populated the form fields for the user to review, challenge and revise in context. The user wasn't chatting beside the product and then copying the result back into the workflow. The conversation and the career goal shared the same surface.
The pattern survived the change in domain.
That makes me wonder how many of the patterns we're currently discovering through coding agents aren't really coding patterns at all.
Humans didn't stop needing review, delegation, clarification or institutional memory as we became better at our jobs. Those structures are part of how we make intelligence useful.
As agents become more capable, perhaps the interesting question isn't only how we make them smarter.