assay · research

The fourteen themes

What the field is actually converging on in agentic engineering, read from the talks rather than the keynotes. Each theme is a short page: the claim, the concepts inside it, how many independent sources back each one, and where Assay honestly stands.

Theme 012 concepts · 20 sources

Context is the new code

The agent stack settled into layers, and the layer humans actually program is context — the rules, specs and skills that surround the model. Everything else is a framework.

Theme 024 concepts · 38 sources

Skills need an SDLC

Reusable context grew faster than anything could govern it. The emerging answer is unglamorous: treat skills as software, with the whole lifecycle that implies.

Theme 036 concepts · 50 sources

Memory and dreaming

Every session learns, and then dies. The theme is what survives the session — memory as an engineered substrate, and learning as an out-of-band pass over what happened.

Theme 0410 concepts · 87 sources

Verification must scale faster than generation

Generation got cheap; trust did not. Coding was never the bottleneck — verification was, and agents multiply it: whatever else scales, verification throughput has to scale faster.

Theme 058 concepts · 73 sources

Isolation beats instruction

A security policy that lives in the prompt is a suggestion. The surviving answers move control out of the prompt and into the substrate.

Theme 069 concepts · 115 sources

The repo is a factory

Continuous AI is arriving as a third pillar beside CI/CD: cadenced agent loops that run inside the repository and treat the pull request as the hard human gate.

Theme 073 concepts · 37 sources

Measure outcomes, not activity

Token spend, session counts and lines written are vanity metrics. What survives scrutiny is whether outcomes landed, what each one cost, and how many nines the fleet holds.

Theme 086 concepts · 55 sources

Humans move up-stack

AI is an amplifier, not a leveler. Goals, systems thinking, constraints and taste stay human — and the talks were unusually specific about what breaks when that is forgotten.

Theme 094 concepts · 38 sources

Agent evals

Evaluation is moving off the pre-launch checkpoint and into production: judge the whole run and its deployment readiness, not just the final answer — and audit the measuring stick itself.

Theme 105 concepts · 29 sources

Agent infrastructure

The substrate under agents is becoming a first-class engineering surface — what agents run on, read from and talk through now gets designed rather than assumed.

Theme 112 concepts · 15 sources

Self-improvement and RL

Getting better without touching the weights: environments, harnesses and memory are both the training signal and the thing being improved.

Theme 122 concepts · 10 sources

World models and embodiment

Learned simulators you can roll a policy forward inside, and the lessons embodied agents keep teaching their software cousins.

Theme 131 concept

Autonomous discovery

Agent systems pointed at the whole research loop — with the human posing the question and accepting what survived.

Theme 142 concepts · 7 sources

The agent economy

Agents as economic actors: an open, inspectable stack underneath them, and payment rails built for principals who are not human.