AI Native DevCon London 2026

We reviewed the full conference playlist — all 41 talks of AI Native DevCon London 2026, worked from the complete transcripts, one written analysis per talk. This section is the distillation: the eight concepts the conference converged on, each with two parts — what it is, and how Assay implements it, including where we are honestly not there yet.

§1The conference in one paragraph

The field has converged on one architecture — spec-first work units, agent loops with narrow permissions, human-gated merges, markdown memory, skills as the knowledge unit — and has moved on to its second-generation problems: skill lifecycle management, memory curation at fleet scale, verification that scales faster than generation, isolation that does not rely on instructions, and measurement that proves value without vanity metrics. Across all 41 talks, nobody advocated removing the human gate; the only divergence was how much evidence the gate is handed.

For Assay this is simultaneously validation and a to-do list. The architecture the speakers described is the one Assay already runs. The problems they are now hitting are, in several cases, problems Assay has designed against and not yet shipped — and we say exactly which ones below.

§2The eight concepts

Each page follows the same shape: the concept and its major points as the conference presented them, then Assay's implementation of it — what is shipped, what is designed, and what is missing.

01Context is the new code assay: match, leads on distribution
The stack settled into layers — models, tools, context, harness, factory lines — and the layer humans actually program is context.
02Skills need an SDLC assay: versioning ahead, evals not yet
Reusable context exploded faster than it could be governed. The fix is software discipline, ported whole: versioning, evals, registries, supply-chain scanning.
03Memory and dreaming assay: substrate shipped, dreaming designed
Markdown memory as a filesystem, plus an out-of-band pass that mines session transcripts into evidence-cited proposals a human accepts or rejects.
04Verification at scale assay: leads
Generation got cheap and trust did not. The spec is the program, the AI is its compiler, and verification throughput must outpace generation throughput.
05Isolation beats instruction assay: identity ahead, no sandbox
A security policy that lives in the prompt is a suggestion. The talks demoed prompt controls failing live and shipped hardware-level answers.
06Outcomes, not activity assay: trails — on the roadmap
Token spend and session counts are vanity metrics. Merge rate, zero-touch merges, and turns-to-correct are the numbers that survive scrutiny.
07The repo is a factory assay: strongest external validation
Continuous AI as a third pillar beside CI/CD: cadenced agent loops in the repository, narrow output channels, and the PR as the hard human gate.
08Humans move up-stack assay: match
AI is an amplifier, not a leveler. Goals, systems thinking, constraints, and taste stay human; the trust layer rebuilds around identity and attribution.

§3One level deeper: the concept register

Each of the eight concepts decomposes further in the concept register — the rolling, per-concept frontier tracker this review seeded. Every new event we ingest (the Agentic AI Summit 2026 was the second) folds its evidence onto these concepts, so the register stays current while this page records how the conference read at the time.

§4What Assay does that the conference did not cover

Eight concepts ran both ways. These parts of Assay had no counterpart in any talk — mechanisms the conference has not arrived at yet.

Table Assay mechanisms with no analogue in the playlist

Mechanism	What it is	Why it was not on stage
Generated, single-writer status board	The board is a build artifact: regenerated from the briefs and registers by a tool that is its only writer. Nobody hand-edits status; there is nothing to hand-edit.	Every speaker measured agent work by mining logs after the fact. None had status itself derived, linted, and single-sourced from the work units.
Registers that resist their authors	Append-only record files with slug IDs, tombstones instead of edits, and deletion detection wired to CI — a deleted record turns the build red.	The talks captured lessons in diaries and distillations, all mutable by design. Durable, tamper-evident memory-as-record was absent.
A spec that discloses its own divergences	The Assay specification is versioned, third-party implementable, and ships with tables of its reference implementation's known deviations — self-declared, itemized.	Speakers shipped practices; nobody published a normative spec, and nobody disclosed where their own tooling fails to meet it.
Release-engineered distribution	Pinned versions, per-platform hashes, release preconditions, and install tooling for the skill bundle — lockfile-grade integrity, not "clone the repo."	Registries and pinning were named as the aspiration. Nobody demonstrated hash-verified, consumer-pinned distribution for a methodology.
Durable cross-machine work claims	Before working, an agent records its claim on a git ref — visible to every machine, recoverable after a crash, releasable when done. Two desks never take the same brief.	Talks had task buckets and voluntary claiming, all single-machine. Claims as first-class, forge-side state did not appear.
Provenance-preserving merges	Two-parent merges only, never rebase; history is the audit trail, so an auditor can walk any line of work back through its review.	Speaker footage celebrated velocity. Provenance of the merged history itself — what an auditor reads — was not discussed.

§5How this review was made

All 41 talks of the conference playlist (event page) were reviewed from their full subtitle transcripts — excluding only a short recap, a promo, two brief interviews, the raw stream, and five unavailable videos. Each talk got a written analysis with key learnings, an explicit what is genuinely new here assessment against prior art, and a relevance analysis for Assay. The eight concepts above are the synthesis of those analyses; every claim on the concept pages traces back to a named talk, and each concept page lists the talks that cover it, linked to the videos.

A note on honesty. Where Assay does not yet implement a concept, these pages say so plainly rather than rounding up. The same rule the specification holds itself to — claim the weaker, true statement — applies to this section.