Ephemeris · Issue 070 Monday · 29 Jun 2026 · Zürich
Ephemeris · Issue 070 01 / 5 · lucumr.pocoo.org
Agentic coding · Essay

The harness runs the loop now.

Armin Ronacher's case: orchestration harnesses — prompt the agent, judge the result, prompt again — are becoming the default way software gets built. They're unbeatable at mechanical work: porting, benchmarking, security scanning. Point one at lasting production code and you get defensive, sprawling systems no human fully holds in their head. The role he sees arriving for the senior engineer is messenger to a machine, not author of the program.

$ while ! task.done; do
    prompt -> act -> judge
  done

# port legacy module ........ ok
# run the benchmark suite ... ok
# ship the prod feature ..... who still reads this?
Ephemeris · Issue 070 02 / 5 · krea.ai
Generative media · Open weights

A from-scratch image model, now open.

Krea published K2 — Raw and Turbo — as open weights, with a full technical report. It's a text-to-image foundation model trained from scratch for aesthetic control and style transfer, and Turbo renders 2K images in roughly two seconds on consumer hardware. The community license clears commercial use for individuals and small teams, which puts a top-ten text-to-image model on your own GPU.

12.9B params
DiT backbone, from scratch
~2sec / 2K
Turbo, consumer hardware
2variants
Raw + Turbo, open weights
Ephemeris · Issue 070 03 / 5 · anthropic.com
Agents · Collaboration

Tag the teammate.

Claude Tag drops one Claude into a Slack channel as a standing member. @Claude a request and it splits the work into stages, runs them against the tools and data an admin has granted, and posts results back in-thread. It carries context across the channel, raises things on its own, and can pursue a multi-day project. Beta for Team and Enterprise.

ScopeOne Claude per channel
MemoryBuilds context over time
InitiativeFlags and follows up
CadenceAsync, multi-day runs
Ephemeris · Issue 070 04 / 5 · openai.com
Security · Open source

Point the model at the code the world runs on.

Patch the Planet, a Daybreak initiative OpenAI built with Trail of Bits, aims cyber-capable models at critical open source to find, validate, and fix vulnerabilities — then routes everything through human review so maintainers aren't buried under machine-generated false positives. A five-day sprint across cURL, the Go project, Python, Sigstore and others surfaced hundreds of issues and merged dozens of patches.

Hundreds surfaced
Dozens merged
Reusable workflows shipped: fuzzing · variant analysis · differential testing
Ephemeris · Issue 070 05 / 5 · cursor.com · via @TochkiNadAI
Dev tools · Agents

Drop a coding agent into the doc.

Notion's External Agents API makes Cursor, Claude Code and Codex first-class collaborators inside a workspace: mention one on a spec, or assign it an issue, and it plans, builds, tests, and opens a PR your team can review. Notion shipped its Cursor integration on the Cursor SDK in a couple of weeks — the SDK handing over sandboxes, model routing and tool use, so the team built the experience rather than the engine.

Notion doc→@agent→plan→build→test→open PR
// External Agents API · runs on the Cursor SDK runtime
End of issue 070 Back to top ↑

That's today.

Today's sources

Armin Ronacher (lucumr), Krea, Anthropic, OpenAI's Daybreak, and Cursor — surfaced via @TochkiNadAI. The recurring thread: agents moving from tools you prompt to teammates that run the loop.

The rubric

Tools a senior engineer or founder could adopt this week — agentic coding, generative media, AI security. Each pick published within the last 30 days. No reprints against the ledger.

Masthead

Ephemeris · Issue 070 · Monday, 29 June 2026 · Zürich. Five picks on a thin day; signal over volume.