Geoff Goodman's TorkBot collapses every message from Slack, GitHub, and Google Workspace into a single LLM conversation, betting that model intelligence will soon handle the cognitive complexity of cross-platform continuity.
That is the controversial core of his octopus architecture. Most agent designs compartmentalize conversations by platform or thread, buying simplicity at the cost of continuity. Goodman instead forces all surface activity through one "foreground lane," then delegates heavy work to semi-autonomous appendages.
Why Collapse All Platforms Into One Conversation?
Three competing pressures drove the design: responsiveness, capability, and continuity. Surface interactions must stay fast, so the foreground lane avoids I/O entirely. Complex tasks require real-time delegation and steering. Continuity demands a single, curated LLM conversation that carries personality and short-term memory as a side effect, not an add-on.
Goodman's bet is explicit: splitting conversations across arbitrary platform-defined boundaries is antithetical to continuity. He wants the agent to seamlessly continue work started in Slack and finished on GitHub, making links across threads and surfaces. "If we’re not there yet in model intelligence, I bet we soon will be," he writes.
How the Octopus Keeps Its Head Available While Arms Work
The architecture uses a central dispatcher (the "brain") that directs multiple lanes (the "appendages"), each with its own context. Lanes can be long-lived static ones like Curator, instantiated templates for bounded tasks, or sandbox VMs with saved filesystem snapshots. Inter-lane communication is just text, betting on prose as the universal carrier of intent.
When the foreground lane needs something done, it picks a lane template and sends an initial message. That lane then owns all the messy work: tool calls, dead-ends, I/O, sandbox workflows. The mess stays contained in the lane's context. Lanes also share artifacts via a virtual filesystem's ./shared folder.
Mean-time-to-interaction is the prize. Completion can take a while, but the foreground must never go dark because one appendage is busy. So the foreground stays small and boring: stable prompt, current intent, recent surface activity, compact summaries and references. Churn stays in the appendages.
The Bet on Emergent Intelligence and Cache Efficiency
A stable foreground prompt means better LLM API cache hits: faster first tokens, less cognitive drag. Curation enables compaction, which keeps lane context from swelling. The Curator promotes durable bits into memory or skills; artifacts remain inspectable without being stuffed back into the foreground.
Goodman admits the cognitive complexity of a single conversation across all surfaces may exceed today's frontier models. But he is designing for the world he believes is coming. If he is right, the next generation of agentic systems will treat platform boundaries as an implementation detail, not a design constraint.
Source: The octopus architecture for AI agents
Domain: blog.goodman.dev
Comments load interactively on the live page.