What does it mean to be an engineer when agents write the software? That was the question behind my short conversation with Kief Morris (Thoughtworks) on the floor at PlatformCon Live Day London 2026, and his answer reframes the entire debate about AIâs role in engineering more usefully than most of the takes I heard that day.
Agents Are the Topic, Not the Problem
Kiefâs opening observation was blunt: agents are the defining topic in the industry right now, but the real challenge was never whether to use them â it is learning to use them well. That distinction matters more than it sounds. Most of the anxiety around agentic development comes from treating âshould we adopt agentsâ as the open question, when the harder and more useful question is âwhat does competent use actually look like.â
Human on the Loop, Not Human in the Loop
Kiefâs article on MartinFowler.com, âHuman on the Loop,â reframes the engineerâs role in a way that survives contact with how agentic development actually works day to day: the job is to build the system that builds the software, not just the software itself.
The distinction between âin the loopâ and âon the loopâ is doing real work here. âIn the loopâ implies a human checking every individual output â a model that does not scale and quietly degrades into rubber-stamping once the volume gets high enough. âOn the loopâ describes someone who has built and continuously tunes the system â the guardrails, the evaluation criteria, the escalation paths â that determines whether the agentâs output is trustworthy in the first place. The engineerâs attention moves up a level, from individual outputs to the system that shapes all of them.
The Talk: âHumans on the Loopâ
The same afternoon Kief took the stage with a session titled âHumans on the Loop: Taking Agentic Engineering End to Endâ. A slide with the session title lists him as Distinguished Engineer, Technology Advisory Services at Thoughtworks, with the cover of his OâReilly book Infrastructure as Code alongside.

The session-title slide from Kief Morrisâs PlatformCon Live Day London talk.
The slides I photographed move through the argument in steps, beginning with âvibe coding (basically)â: an idea becomes a prompt, the agent builds, working software comes out, and a human reviews it.

âVibe coding (basically)â: prompt, build, working software, review.
The next slide adds a spec at the start and tests at the end: âSpec-driven development with testsâ puts the spec and the tests around the build step, so the review loop has something concrete to check against.

âSpec-driven development with testsâ: the same loop with a spec in front and tests behind.
The ideas also connect with the âagentic coding harnessâ slide from the same deck, defined as âa system that drives the LLM to make good codeâ. That is the âsystem that builds the softwareâ in practice.
Sensing Deliverability and Operability
The slides also cover how a human on the loop knows whether the output is any good. One slide asks âHow do we sense deliverability?â, with DORA metrics as the lagging sensors and mutation testing and fuzz testing as leading sensors.

âHow do we sense deliverability?â with DORA as the lagging sensor.
A follow-up slide asks the same for operability and splits the answer into lagging sensors (DORA deliverability metrics, user satisfaction, business operations metrics such as transactions and cost, commercial metrics) and leading sensors (deployability, performance and scalability testing, failure scenario testing, penetration testing, compliance auditing).

âHow do we sense operability?â: lagging sensors on the left, leading sensors on the right.
My take: this lines up with the âwhat does not changeâ point from our conversation. Quality and the path to production still need measuring, and the leading-sensor list reads like the checks a platform team can automate in the pipeline.
What Does Not Change
The part of the conversation I keep coming back to: managing quality and the path to production remain paramount, exactly as they always have. AI changes the tooling â what generates the first draft of the code, what catches an obvious mistake, what accelerates a refactor â but it does not change the fundamentals of what makes software fit to ship. A codebase with weak test coverage, unclear ownership, or no deployment discipline does not get healthier because an agent wrote more of it faster.
This is the same conclusion I keep arriving at from a different angle in production guardrails for AI agents: the discipline lives in the system around the model, not in the model itself.
Find His Work
Kief pointed me to infrastructure-as-code.com for the deeper material â or, as he put it, just search âKiefâ (K-I-E-F) and the right results surface. Given how much of the current agentic-AI conversation is repeating lessons Infrastructure as Code already learned about treating operational systems as engineered artifacts, that back catalog is worth the search.
Related Reading
- PlatformCon London 2026: The AI Era Runs on Platforms
- Your IDP Is the Foundation for Agentic AI
- Production Guardrails for AI Agents
About the Author
I am Luca Berton, AI and Cloud Advisor. I work at the intersection of platform engineering, cloud security, and enterprise AI deployments. Book a consultation.


