A minimal agent harness written in Bend2, built for the Bend2 orchestrator (baton; the orchestrator may be renamed Orchestra). The harness runs one agent seat: it receives tasking, runs model turns against a small tool set, and reports back through the orchestrator's message protocol.
One seat, one loop: read the assigned task, run a model turn, execute the tool calls the model requests, append the results to the session, repeat until the task reports completion. The loop, the tool dispatch, and the session format are written in Bend2.
- The turn loop, tool dispatch, and session persistence are written in Bend2.
- The tool set is minimal: read files, write files, edit files, run shell commands, and read the orchestrator inbox.
- Tool calls run asynchronously. A tool streams its output into the session as the tool produces it.
- The harness speaks the orchestrator's existing CLI protocol (baton2) for task assignment, messages, receipts, and completion requests.
- One session maps to one agent seat. The session format records turns, tool calls, and tool outputs as discrete events so the orchestrator can resume a seat at any turn boundary.
Each project below is prior art worth studying for a specific set of decisions. The links are the canonical repositories.
- Pi (Mario Zechner) — a minimal coding-agent harness decomposed into small packages: a multi-provider model gateway, a stateful agent-loop core, a terminal UI, and the CLI that binds them. Study its package boundaries and its small default tool set (read, write, edit, shell).
- DeepSeek Harness (
dsh) — separates model weights from harness scaffolding and defines the harness as loop, tools, session storage, permissions, and security boundaries. Its named runtime modes (standard, programmatic tool calls, minimal) are worth studying, especially the minimal mode: a file editor, a shell, and a compact system prompt. - fx (Vercel Labs) — a coding-agent CLI implemented in Zig and shipped as a single native binary that streams conversation output into terminal scrollback and delegates work to subagents. Study its single-binary footprint, its native tool primitives, and its use of the terminal as a plain Unix stream.
- Prime Agent (Prime Intellect) — a long-horizon harness built around a persistent code kernel: context is a variable the model can read and rewrite, subagents spawn programmatically from inside that kernel, and durable agent state (memories, skills, prompts) is updated through small evidence-backed edits. Study its recursive-subagent model and its continual-session design.
- What is the minimal tool set that keeps an agent effective on the orchestrator's task profile?
- What does asynchronous, streaming tool execution require from the session format and the turn loop: ordering of concurrent results, chunked output events, cancellation of in-flight tools when a turn ends, and completion tracking for long-running remote gates?
- How should programmatic tool calls (generated code that batches tool operations) compare with discrete tool calls on this workload, and where is the crossover?
- What Bend2-native session format records turns, tool calls, and outputs so the orchestrator can resume or fork a seat at a turn boundary, and compact a seat's history without losing tasking?
- What does a minimal permission and sandbox model look like in Bend2, given that all compiler and test gates already run remotely on designated build hosts?
Design phase. No implementation yet. The first milestone is a Bend2 session format and a turn loop that can complete one orchestrator-assigned task end to end.