Trinity v0.8.5 — Agents That Coordinate

Trinity v0.8.5 is out.773 commits, ~95 issues. The theme: agents that coordinate — with each other, and with time.

Read the full write-up in What's New →

Highlights

• Task-completion events— agents subscribe to each other's task terminals and get woken with the outcome the moment work finishes. Orchestrators stop polling; long-running handoffs just work.
• Agent self-reminders— a running agent can schedule a future re-invocation of itself (“check this again in 2 hours”). Durable, survives restarts, visible in the execution timeline.
• Shared rooms— multi-agent sessions: several agents in one conversation with a mention-wake turn engine, budgets, and a live Sessions view on the dashboard (enterprise).
• Skills are full packages now— a skill is a directory (scripts, templates, resources), versioned and safely synced; plus a stateless skill-runner that executes library skills over MCP (enterprise).
• Display names— rename the label, never the slug: friendly agent names across the whole UI without breaking URLs, automations, or integrations.
• Ghost agents— disposable, budgeted agents for burst work that hard-discard themselves at their execution/TTL budget (enterprise).
• Bring your own GitHub token— per-user PATs (each user's own repo scope), and public github: templates now clone with no token at all.

Plus voice replies v2 (voice is now a per-message agent choice), local-first telemetry (default-on, never leaves the box), a big fleet-hygiene pass (git bloat root-caused, Docker volume reclaim), and a blast-radius guard on data retention.

What's new → docs.ability.ai/whats-new/v0.8.5

Full notes → github.com/abilityai/trinity/releases/tag/v0.8.5

Discuss on GitHub → github.com/Abilityai/trinity/discussions/1858