The Fourth Institute ships as v1.0.0: one browser workspace runs a research team of agents from different AI providers
Synopsis
The Fourth Institute has been released as v1.0.0, an open-source workspace that puts agents from different AI providers, including Codex, Claude Code, DeepSeek Harness, OpenCode and Kimi Code, behind one browser interface so a single person can organize agent teams, split work across backends and keep research running. Persistent residents, lineages and families organize members and task handoffs, and agents reach resources through the HATG ISA instruction set; the team reports that a team organized by one person recorded over 2 billion tokens of usage in a single hour under high load, while all project content is AI-generated and the bundled cellular-automata study is a demonstration with no academic value.
Interpretation
Agents from different vendors can now be directed, teamed and monitored from one browser interface, instead of each being driven through its own separate terminal program. Previously such agents were typically operated through separate terminal applications, with conversations, accounts, execution status and usage scattered across them. The release is v1.0.0, and the interface covers Codex, Claude Code, DeepSeek Harness, OpenCode and Kimi Code, and can connect cloud or self-hosted endpoints speaking OpenAI Chat Completions, OpenAI Responses or Anthropic Messages.
Collaboration rests on three organizing units: residents give agents persistent identities, workspaces and records across sessions, lineages connect members' work through events and task handoffs, and families give closely collaborating members a shared entry point. Cross-backend collaboration only works if members can share resources and exchange results rather than each answering alone. Agents use HATG ISA, the Human-Agent Trajectory Graph instruction set, to access resources and coordinate work, and shared papers, knowledge entries, code and workspaces carry findings into the next task; the example split assigns implementation to a Codex agent, review to a Claude Code agent and experimental analysis to a DeepSeek Harness agent.
Protocol conversion decouples the execution environment from the model vendor, so Claude Code can use an OpenAI-only API and Codex can use an Anthropic-only API. A native protocol should be preferred when available, and conversion is used only when the service does not support the chosen harness's protocol. Conversion runs locally and carries text, images, function calls and results, streamed output, usage and reasoning state; it requires full conversation history and harness-owned tools, and vendor server tools and server-side Responses conversation and compaction IDs cannot cross protocols; when no explicit effort is supplied, thinking token budgets are approximated as low (under 4,096), medium (under 16,384) or high.
Perspective
The demonstrated scope is a v1.0.0 open-source workspace: operating multiple agents from one browser, organizing teams by persistent residents and lineages, converting protocols locally, and letting humans and external agents join projects within a station. The bundled cellular-automata project serves only as a workflow demonstration, and its papers and experimental findings have no academic value; public agent uploads, discovery across stations and automated collaboration between stations remain future directions.
The team reports that a team organized by one person recorded over 2 billion tokens of usage in a single hour under high load, which is an experiential description rather than a controlled measurement; long-running teams can quickly fill disk, and storage management, deduplication and cost control are listed as areas for further optimization; the project states that all content is AI-generated, unexpected problems may arise during installation or use, and scientific reliability and research value still require validation.
