v0.1.68 — Guided agents actually run (and leave real files behind)

Released 2026-07-28. Signed + notarized on Mac (Apple Silicon + Intel); Windows + Linux published, Windows unsigned pending Azure. Gated on full smoke + both E2E journeys.

Build an agent by describing it — and it just runs

When you build an agent the easy way — describe it in plain language and hit Build my agent — two things are now true that weren’t before:

• You get real files. The guided builder used to write only a hidden manifest, so your project folder looked empty (ls showed nothing) and there was nothing to inspect, fork, or hand to an engineer. It now scaffolds a real agent.py alongside the plan, so the agent is actual code you can open, edit, and graduate.
• It runs with no API key. Testing a described agent used to fail with “No agent.ts or agent.py found at project root” — because a described agent is a plan, not a key-carrying code process, and the card tried to run it the wrong way. Now the card’s Send routes it through the keyless AgentFoundry engine (the same one Business-mode Build uses), so it runs on our served inference with no key to paste, and the result shows up right in the card.

Bring-your-own-key agents (the “build from a kit” path) still run your agent.py directly, exactly as before.

Full smoke green before tag (static 0/0; runtime 258 pass; both journeys 24/24 + 37/37). The runtime failures are the known sandbox pty false-positives that pass on CI.