Where things stand. The Alpha (v0.1.0-alpha) is what ships today: a single-machine proof that the multi-agent SDLC loop works. A VS Code extension plus a local workflow server run specialized personas for every phase — requirements through deployment — with a task board, artifact linking, semantic search over the project, and a first-pass signoff and quality-gate pipeline. The loop runs, and it’s free to install.
I’d rather you hear the limits from me: in the Alpha, governance is light, work is effectively one agent at a time, and the record underneath is markdown plus search. The right mental model is still a capable new teammate who needs direction, not a hands-off autopilot.
What Beta changes. Beta is working — I run it daily. The control plane is in place; I’m tuning the governance and working through bugs, and it ships when it clears validation; I’m not going to pin a date on it. The difference comes down to shifts in both capability and structure.
1. Governance becomes the architecture. The Alpha records signoffs and runs basic gates. In Beta, signoffs carry evidence and can be required, issued, or revoked; readiness, scope, and completion checks gate every stage; acceptance-criteria coverage and testing evidence gate the handoffs; and when an agent trips a rule, the system responds with corrective guidance. Governance stops being ad-hoc checks and becomes evidence-gated process at every stage of the lifecycle.
2. Parallel agents on one shared codebase. The Alpha works one agent at a time. In Beta, multiple coding personas work simultaneously without clobbering each other — per-file locking with a check-in lifecycle, a pull board of code-ready work, and the endurance to keep long-running work on the rails.
3. A knowledge graph as the record — and you can watch the work. In Beta, the source of truth moves from markdown-plus-search to a knowledge graph built for traceability — tracing a requirement through to its tests becomes a query rather than a memory exercise. And the work becomes observable: you can watch what each persona session is doing.
4. Always-on reliability. In Beta, the local server runs supervised — liveness checks, a watchdog, safe restarts — so the agents’ tools stay reachable at all times rather than depending on a fragile local process.
5. A governed operator surface. In Beta, the task board grows into a governed workflow view: a full task lifecycle from icebox to closed, board filters, work-in-progress visibility, and richer task management.
Further out. The roadmap carries cloud and multi-user collaboration as the next structural layer beyond Beta.
The short version: the Alpha proves the multi-agent SDLC loop works locally, for one user, with light guardrails. Beta makes it governed — enforced and evidence-gated — concurrent, graph-backed and observable, and reliably always-on.
Today’s build is fully functional; if you’re evaluating, grab the update when it lands — and tell me where it breaks.
Crucible is an AI software-development platform from Ember Agentic Labs that builds working, tested software by directing specialized AI agents — the alpha is free to install — crucible.emberagenticlabs.com. Built by Tony Gibbons.