The Code Writes Itself. The Judgment Doesn't.
What AI Native DevCon 2026 revealed about the real bottleneck in agentic development
Five hundred builders spent two days at The Brewery in London (June 1–2) asking an uncomfortable question: what is software development, once agents do most of the typing?
The conventional story of the past eighteen months feels like success. The agent scaffolds a feature in minutes. The demo works. Leadership is impressed. Teams report individual productivity gains that would have sounded absurd in 2023. And that is exactly where most organizations have stopped.
Guy Podjarny, the Tessl founder who organizes the conference, framed the turn: 2025 was the year coding agents showed real promise; 2026 is the year we find out whether they hold up in production — across teams, codebases, and environments, without constant human correction. The useful question is no longer "can the agent do it?" It is "can we govern what the agent does, at the speed it does it?"
The bottleneck has moved from writing code to governing it — and the industry is now rebuilding the entire software stack around a new unit: the skill.
Four themes ran through the two days.
Skills are becoming the unit of software — with none of the infrastructure. A skill is a reusable, versioned set of instructions for an agent, and the thesis threading the conference was that these artifacts, not source files, are becoming what teams author, share, and depend on. The problem: we have recreated the early days of programming without any of its safety net. There is no mature static analysis for skills, no testing discipline (evals are the embryonic equivalent), no dependency management, no observability. Snyk's Liran Tal made the gap concrete in the best-titled talk of the event: "Your agent installed malware because a SKILL.md told it to." A skill is executable trust. Right now, almost nobody audits it.
Harness engineering is context delivery. OpenAI's Ryan Lopopolo argued that the binding constraints of software development have changed: they are now human time, human and model attention, and the context window. His prescription was to stop hoping the model infers your non-functional requirements and instead encode them into the harness — the review surfaces, approval gates, and context pipelines around the agent. Think of the harness as a factory jig: the craftsman's judgment, cast into tooling so every run inherits it.
Agents don't learn, so memory becomes architecture. Anthropic's Lamis Mukta named the quiet flaw in the agentic dream: intelligence doesn't compound. Task fifty starts as ignorant as task one. The industry's answer is a progression — from static instruction files to memory tools to skills to agent-managed memory, including out-of-band consolidation processes she described as "dreaming." The sharpest audience question of the conference cut through it: at what point are we reinventing databases from first principles? The honest answer seemed to be: we already are, and we should at least do it deliberately.
The human is now the congestion point. Thoughtworks' Birgitta Böckeler closed the event by naming the flow crisis: agents generate code faster than humans can review it, and the real costs sit beyond tokens — in the "harness tax" each tool imposes and in human energy. GitHub Next's Don Syme offered the industry a useful polarity: the hype lives in individual productivity, but the unsolved problem is team and SDLC continuity — what GitHub is betting on with agentic workflows and "continuous AI." Netlify's Dana Lawson added the design lens: platforms now need AX, agent experience, alongside UX and DX, because half the users of your APIs will soon be non-human. The scarce resource, she argued, is no longer typing speed. It is taste, judgment, and architecture.
The most grounded proof that this is not conference-circuit theorizing came from ReCinq and Odevo, a 14,000-person property management company restructuring itself as AI-native — with Meta's Ian Thomas describing the same reorganization at engineering-org scale.
For a team lead, the diagnostic writes itself. How many of the instructions your agents follow have been reviewed the way you review code? Can you name who authored the skills in your pipeline — and would you notice if one changed? When your agents double their output next quarter, what happens to your review queue? If any of these answers make you wince, the conference's message was for you.
Getting an agent to work is a demo. Getting a thousand runs to agree is engineering.
Based on AI Native DevCon 2026 (Tessl), The Brewery, London, June 1–2, 2026. All 40+ talks are available on demand at tessl.io/devcon.