The Agent Multiplexer Is Becoming a Runtime

Herdr 0.8.2 is an open-source terminal multiplexer built around coding agents. It owns real PTYs, keeps sessions alive, tracks agent state, and lets people and agents act on the same server-owned panes through client and automation protocols. That shared session is enough to call Herdr an early agent runtime. Its current limit is also in that path: much of the lifecycle still comes from interpreting a terminal screen, and its waits track state rather than one turn.

Mitchell Hashimoto announced Superlogical on July 29. The company will start with a terminal multiplexer, then make its parts composable and safe for production. Herdr already ships a narrower version of that bet. Herdr uses libghostty; Superlogical says it will build on the same public components. Superlogical is a rival theory, not yet a code-level competitor.

The pane became an addressable agent

Mature multiplexers already keep sessions alive and expose automation surfaces. Herdr packages a more specific contract: every agent gets a stable name, state, terminal identity, and command surface. A person can watch the TUI while another agent runs herdr agent prompt reviewer "Check the diff" --wait. Both use the same server-owned session.

The server protocol is not a thin wrapper over keystrokes. Protocol version 20 frames requests over local sockets, caps normal payloads at 2 MB, and negotiates either semantic frames or terminal ANSI for rendering. The server owns the workspaces, tabs, panes, and PTYs. Clients come and go.

An agent multiplexer becomes a runtime when control and lifecycle share the same durable session.

That claim can be proven wrong. If state is too weak for automation, or recovery restores a layout without restoring the work, Herdr remains a good multiplexer with agent labels.

A prompt is one server operation

1. Resolve Agent target

   Pin pane and terminal identity
   validate →
2. Submit PTY input

   Write text, then encoded Enter
   observe →
3. Track Lifecycle event

   Gate activity from a non-working start
   settle →
4. Return Agent state

   Idle, done, or blocked

The request handler first rejects an empty prompt. It resolves the named agent to a pane and checks that the pane still contains the expected foreground process. If the agent is already blocked, Herdr returns agent_blocked before it writes anything. Otherwise it encodes the prompt for the pane's live terminal mode, sends the text, waits 300 milliseconds, then sends Enter.

The wait path records the event sequence and terminal identity before submission, so a replacement process cannot satisfy the old wait. From a non-working start, Herdr caps its activity gate at five seconds or the caller's shorter timeout. Only a longer caller timeout can produce agent_prompt_stalled. It then waits for idle, done, or blocked by default.

State has two authorities

Herdr can learn agent state from a native integration or from the terminal itself. Integrations call pane.report_agent with a source, sequence, state, and optional native session reference. Session-only integrations report an ID without changing waits or status. When no hook owns the state, Herdr inspects the foreground process and visible terminal shape.

Version 0.8.2 shows the cost. Its fixes teach Qwen Code's localized confirmation states and strip new Claude Code title-spinner frames. Herdr can support a terminal agent without owning its API, but it must follow that agent's UI changes.

The detector damps that noise. A transition from working to plain idle waits for repeated confirmation, and stable blocked signals refresh on a timer. The project docs still say an unfamiliar prompt can appear idle instead of blocked until Herdr learns its screen shape. That mistake changes the sidebar and waits. It does not make Herdr answer the prompt by itself.

Persistence has four meanings

“The session survives” is too vague for an agent runtime. A client can detach while the server and PTY keep running. A restarted server can restore workspace and pane state. Supported integrations can resume native agent conversations from stored IDs. Experimental handoff tries to transfer live panes to a new server during an update. Those are separate guarantees.

Herdr keeps the separation in code. persist.rs stores session state, optional screen history, and plugins in different files. The native-resume layer accepts session references only from recognized Herdr integration sources, limits their size, and builds an agent-specific resume command. A saved screen and a resumed Codex conversation are not the same artifact.

I inspected these paths but did not run the recovery test for this issue. The adoption verdict stays inside that evidence boundary: detach is a core product path, while restart and handoff need a controlled trial before any of them carry unattended work that matters.

Superlogical is the rival theory

Superlogical calls the missing layer a durable session around the work itself. Its three-step plan starts with a terminal multiplexer, makes its parts composable, then makes the system safe for production. The first product promises web, macOS, and iOS clients, live sharing, and native scrollback and selection.

As of August 20, no public Superlogical multiplexer repository or beta was linked from its announcement or site. Its architecture is not available for a code comparison. The public claim I can test is narrower: it will use the same MIT-licensed libghostty components that Herdr already vendors.

libghostty-vt's public terminal API accepts terminal output, maintains screen state, and exposes it to an embedder. It does not create or manage the PTY. The public C API is also marked incomplete and subject to breaking changes, which explains why Herdr pins a source commit. Superlogical must still build process supervision, durable sessions, transport, auth, sharing, storage, and client policy.

libghostty lowers the cost of terminal correctness. It does not ship the multiplexer.

Herdr and Superlogical may fit together. Herdr is opinionated about coding agents and works inside an existing terminal over SSH. Superlogical says agents are one kind of work inside a wider session, with native and web clients from the start. They still compete for the durable terminal session. Partnership remains a theory until either project publishes an integration.

Use the narrowest owner that solves the job

Try Herdr now when several CLI agents run on a local or remote machine and both a person and another agent need to inspect them. Keep the multiplexer you already trust when general terminal persistence matters more than native agent state. Watch Superlogical when live sharing, native clients, or production policy is the job, but wait for a beta before treating it as an alternative. Herdr has crossed the line from multiplexer to early runtime, but reliable turn identity and recovery will decide whether it stays there.