While Brian slept, one of his AI agents shipped a blog post, a workout app, and a darts scorekeeper — on its own machine, its own GitHub, and its own AWS bill.
That agent is named Langa Stina, and she's one half of the story in Brian Bischoff's Usable Fragments talk, "Two Operators, One Memory." Brian is CTO of Globe Tracker and Founder & Chairman of Usable, and he runs two autonomous agents — built on the open-source Hermes Agent — that share one Usable memory layer.
Two operators, two very different jobs
Langa Stina is the personal operator, and she runs with the handbrake off. Brian gave her her own machine, her own Gmail, her own GitHub, and her own (sub-)AWS account, with credentials in a 1Password vault. The guardrails are deliberately soft: conventional commits, private-by-default repos, push when done, and — the important one — write everything back to Usable. "Because the blast radius is mine, I let her run." She's shipped around ten projects on her own, including the very plugin that wires Usable in as Hermes' memory.
The Globe Tracker Assistant is the production operator, and it's a different animal entirely. Same engine, same memory layer, same Slack interface — but a tight harness. It works on a locked-down workstation with no elevated access, a GitHub account that can open pull requests but never merge to main, and reaches the rest of the world only through carefully chosen MCP servers. It runs daily repo audits against company IT policy, picks up Jira tickets, opens manifest PRs through the GitOps pipeline, runs smoke tests, and helps triage production incidents — every action branch-protected and review-gated.
Scope by blast radius, not capability
That contrast is the whole point. The engine is identical; what changes is the plane it operates on. A personal project where the only thing at risk is your own weekend gets free rein. Production work that affects customers gets branch protection, review gates, and observability. Same agent, different harness — scoped by blast radius, not by capability.
Memory first
Underneath both sits one idea Brian keeps coming back to: memory is load-bearing. Usable is the persistent brain; MCP is how the agent reaches into the systems where work happens; Slack is the cockpit where a human stays in the loop. Get the order wrong and it falls apart — as Brian's agent put it:
"Memory without MCP is a diary. MCP without memory is a forgetful intern with admin credentials."
The magic isn't the autonomy. It's that both operators write everything back to the same memory — so the work compounds, and a human can always ask Usable what either of them actually did.
Watch the full talk
Brian walks through the arc from typeahead to full autonomy, why he chose Hermes, how he prompts both agents with a single Slack message, and a refreshingly candid Q&A on cost and the runaway-spend lessons that got him there. It's the second of four talks from Usable Fragments 2026.
Usable is knowledge management for AI agents — the shared memory layer behind both of Brian's operators → usable.dev
Watch the full series → Usable Fragments 2026 on YouTube