Blog

Notes from the team building a4sx: what we learn about agent sessions, models and open source along the way.

• i tried to make my agents self-organize by giving them sleep, hormones and peptides - part 2 · 2026-10-23
  Evolving a self-organising agent mesh in a simulator, then checking it against real agents.
• Our agent mesh tied a single agent. So we're teaching it to organise itself. · 2026-10-22
  Real sessions in the Claude Code harness on 52 bug fixes, a workload where one warm agent forgets, a self-organising mesh in simulation, and first tests of each piece on real sessions.
• One agent can't hold a whole platform. So we're building a mesh that learns who knows what. · 2026-10-21
  Where our agent-mesh work stands: a local mesh measured from 2 to 128 agents, a knowledge-mesh design inspired by memory, and a first simulation with scripted agents.
• Germany's AI asset is its knowledge. "Open your data" won't unlock it. · 2026-10-20
  Jensen Huang is right that Germany's knowledge is its most valuable AI resource. The Chancellor's appeal to companies to open their data won't unlock it. Here is why, and what can.
• Is git our bottleneck? Anthropic's source-control post, read from a small agent shop · 2026-10-20
  Anthropic moved off git for agent-scale development. Our numbers say our bottleneck is landing, not git itself, and that we can fix it on git today.
• China's micro-drama boom is already trading prompt know-how · 2026-10-14
  In China, AI micro-dramas spawned a market in courses and exact prompts. What the numbers support, what they don't, and why the full session is worth more than the last prompt.
• What it takes for a small German company to sell software outside Germany (and why we said no to most of the world, for now) · 2026-10-14
  We tried to open a4sx beyond the EU. VAT regimes, US state nexus, sanctions, Stripe's limits and Terms churn followed. What we hit, what we decided, and why most of the world stays closed.
• Agentic genetic programming needs a world to fail in · 2026-10-12
  A model proposes the change, a fitness signal keeps the winner. Such a loop is only as good as the world that scores it, and a producing organisation needs a simulated one first.
• Your agent sessions are worth something: publish and sell them on a4sx · 2026-10-12
  Publishing and selling sessions on a4sx is now free for every signed-in account with a verified email address. Why a session has value, when one is worth selling, and what it costs.
• Novel, recalled or borrowed? OpenAI's math results and the missing transcripts · 2026-10-09
  OpenAI published hundreds of AI-made math manuscripts. Without full transcripts, outsiders can't tell new reasoning from recall, or rule out other sources.
• Everything becomes anything: why we're betting on users, not IP · 2026-10-06
  Atrioc's "Everything Is Everything Now" argues there is no moat on software. Why we think IP is inflating, and what we are doing about it.
• Weights on your own disk, and the awkward question of distillation · 2026-10-06
  Open-weight models are becoming the default for serious builders. What Fireship's Code Report on Ajax shows about data, distillation and provenance.