Observability and multi-provider agents. This release makes what your autonomous agents do — and what they cost — visible, and lets specialists run on more than one model.
Observability
- Run telemetry —
af telemetry on|off|status|report|usagerecords per-step latency and token usage for every agent and formula instance: a local timing table (af telemetry report) and a backend token/session query (af telemetry usage). Off by default; opt-in factory-wide, taking effect at the next session launch. - Web console Telemetry view — the optional loopback console gains a Telemetry tab: per-step timing (Duration), per-run token usage, and session metrics, with a banner that reports backend degradation as data rather than hiding it (when a backend can't attribute tokens to a single step, the console says so instead of inventing a number).
Multi-provider agents
- Agents can run on non-Claude models:
af config models(show/set/check/attest) over amodels.jsonregistry,af install --agents --litellmto stand up an OpenAI gateway,af sling --modelto override per launch, and a fitness-attestation gate before a non-default model flies. Shipsgpt-fable-reviewandgpt-rootcause-all— review and root-cause formulas that run on OpenAI models.
Reliability & safety
- Operator-only factory teardown — factory-wide
af downis now gated so an agent cannot tear down the whole floor. - Reliable improvement self-edits — the AND-gated continuous-improvement hook (
af improvement) hardened so agents apply post-run formula edits reliably.
Formula & agent library
- New shipped formulas:
fable-secure(security-program review),multi-agent(multi-perspective architecture consultation), andmarketing-cycle(a self-marketing cycle for the repository the factory serves). Twenty-four formulas now ship.
Docs
- README refreshed: an Observability section, Web Console + Telemetry coverage, multi-provider notes, and the command reference brought up to date; corrected the formula count (24) and skills list (10). CHANGELOG updated.
Full changelog: v0.1.0...v0.2.0