openclaw - 💡(How to fix) Fix Feature Request: Tenacity — Goal Completion Without Intermediary Confirmation Handoffs [2 comments, 2 participants]

Problem

Hermes Agent (Nous Research) implements a "Tenacity" feature that keeps the agent running continuously until a goal is reached, without stopping to ask for confirmation at each step. The agent only asks for human input when it genuinely cannot proceed (blocked, needs a decision, or needs information).

OpenClaw is gateway-first and turn-based — it operates in discrete confirmation cycles. This is great for safety and control, but limits autonomous execution on long-running tasks where the path is clear.

What I am proposing

Add an optional Tenacity execution mode to OpenClaw that:

1. Goal-oriented loop — Agent keeps executing until the stated goal is reached, or until it hits a hard blocker that genuinely requires human input
2. Checkpoint / auto-save — If the session is interrupted (timeout, crash, reconnect), the agent resumes from the last completed step, not from scratch
3. Heartbeat + zombie detection — If the agent loop stalls, detect it and either restart from checkpoint or alert
4. Session auto-resume — After a disconnect, the agent picks up where it left off
5. Kanban-style progress tracking — The agent reports progress milestones internally so the human can see where it is in the flow

Why this matters

• Long-running tasks (data processing, multi-file refactors, research compilations) are currently fragile — a timeout in the middle means starting over
• Hermes has this and it is cited as the #1 reason power users prefer it for autonomous coding
• It would complement, not replace, the current turn-based model — users who want safety can keep it, users who want autonomy can enable Tenacity

References

• Hermes Agent v0.13 "Tenacity" feature: https://blakecrosley.com/guides/hermes
• Closed learning loop concept: Hermes stores durable identity + persistent memory on disk, aggregates skills that improve with use
• Session auto-resume + Checkpoints v2 are part of the same feature set

Implementation suggestion

This could be implemented as:

1. A new session mode / flag (openclaw run --tenacity)
2. A new plugin type that wraps the agent loop with checkpoint logic
3. A skill that instructs the agent to behave more autonomously when explicitly tasked with a long-running goal

Happy to help test or prototype if the team is interested.