Coding agents · human-in-the-loop

Stop Babysitting Your Coding Agents

Stay informed. Answer when they ask. Steer when needed. Keep the work moving from the conversation.

macOS · Linux · Windows · Open Source
[Lz] NightMe
You · on the move
/steer use the existing parser — no new dependency.
NightMe Agent
Steering. Will switch to nom-free path after the current edit.
🤖: claude · MiniMax-M3
💰: “ 5.1k / 50.9k · $5.14 ”
📁: nightmedev/website · 🌿 guide-coding-agents · #142
⌨️ Working
The conversation itself is the status surface.
What babysitting looks like

Babysitting is the loop you fall into.

A coding agent keeps working. You keep checking. You answer a question, redirect a wrong turn, and decide what comes next — then the loop repeats. That repeated checking, answering, redirecting, and planning is babysitting.

Definition

Babysitting a coding agent means repeatedly checking, answering, redirecting, or planning work just to keep the agent-driven workflow moving.

The human-agent loop
Observe Respond Direct Continue
agent keeps working
Observe

What is the agent doing right now — and is the direction still right?

Respond

Answer a question, grant a permission, or clarify the request when the agent asks.

Direct

Stop a wrong turn or steer a new direction when the work drifts.

Continue

Queue the next instruction, switch project, switch agent, or hand the work off.

Proactive status — from the conversation

The conversation itself is the status surface.

NightMe does not give you another dashboard to watch. The agent’s state travels with the conversation — through IM message state, heartbeat / activity signals, and reaction emoji. You glance at the chat, you know what is happening.

Real Telegram screenshot of the NightMe bot replying to /gtw pr: a light-blue message header reads Devin /gtw pr; the reaction row carries a green checkmark, a cloud with the count 6, a wrench with the count 2, and a clock reading 08:26:50. The body shows PR opened with branch fix-telegram-md against main, a clickable URL https://github.com/cnlangzi/nightme/pull/398, a worktree path /Users/geax/code/geax/github.com/cnlangzi/nightme.nightme/fix-telegram-md, and a risk line reading low — documentation-only change; no code, API, wire, or runtime behavior is affected. A statusbar at the bottom lists agent pi running model senenova-6.8-flash-lite at commit 01a0a79b-854f-734e-90a6-651fb81cdc32, token usage 6.9k / 19.5k / 1.5k, and the folder nightme.nightme/fix-telegram-md on branch fix-telegram-md at PR #398. A small timestamp 08:26 sits in the bottom-right corner.
/gtw pr — real Telegram evidence

Same status surface, real channel: reaction counts (cloud 6, wrench 2), the 08:26 timestamp, and the full /gtw pr output — PR, branch, base, URL, worktree, risk, agent, model, tokens — delivered to a developer’s phone.

The loop, in conversation

Four phases, one conversation.

Every phase of the loop happens inside the same chat. A few short commands cover the cases where you actually need to intervene; everything else is the conversation reacting to what the agent is doing.

Observe

Read the state signals.

The statusbar on each bot message carries agent, tokens, git, and cwd. Reactions and message state tell you whether the agent is working, waiting, or done. You decide whether to keep moving or look closer.

Respond

Answer in the chat.

When the agent asks, reply in the same conversation. Permission prompts, clarifications, and follow-ups reach you where you already are — one reply, then it keeps going.

Direct

Step in when the direction changes.

If the work drifts, /stop cancels the current step and parks the session. /steer <message> appends a new instruction to the agent’s queue — the current step finishes, the new one runs next. Sending any new message after a stop automatically resumes the session.

Continue

What is next.

Queue the next instruction with /steer, switch project with /cwd <path>, or switch agent with /use <agent>. The conversation stays open; the context changes underneath.

A realistic sequence
nightme — one session, four phases
Observe
// statusbar: claude · 5.1k tokens · guide-coding-agents · working
Respond
You grant the bash permission · continue
Direct
You/steer skip the new dep, reuse the existing parser
Continue
You/gtw pr ship this branch as a PR
Multiple work threads

Conversations map naturally to projects and work.

Each conversation owns one project and one agent session. Switch by sending a message in a different chat — not by juggling terminal windows.

Conversation
Project / Workspace
Agent + Context
Current work
Next work

Three chats, three agents, three projects — the same NightMe workflow. See the parallel walkthrough →

No new dedicated app

Use the chat experience you already know.

NightMe connects to supported IM apps and routes every message to the local coding agent already running on your machine. Two non-lock-in properties: your agent can change, and you do not need a new communication app.

Your familiar chat app
a supported IM you already know
NightMe
conversation & control layer
Local coding agent
any coding agent you already run

The agent does the work. NightMe keeps the human-agent loop within reach.

Away from the terminal

Stepping away is a side effect of the same workflow.

Commuting, in a meeting, on another machine — these are not the point. They fall out of the workflow naturally once the conversation itself is the control plane. NightMe does not schedule the work, decide when it runs, or guarantee overnight completion; the underlying local agent keeps doing what it already does, and you stay informed through the chat.

The agent keeps working. You remain connected without babysitting it.

Where NightMe fits

The agent does the work. NightMe keeps you connected to it.

NightMe doesn’t replace your coding agent. It connects the coding agent already running on your machine to the chat experience you already use.

  • 1
    Proactive conversation-based status signals

    IM message state, heartbeat / activity, and reaction emoji — so the conversation tells you what the agent is doing without polling the terminal.

  • 2
    Familiar IM interaction

    Reply in the chat you already use. No new proprietary app to learn or migrate to.

  • 3
    Direct human intervention when needed

    Answer, redirect, or stop — the conversation is the control plane, and step-in is one message away.

  • 4
    Agent-agnostic, workflow stays the same

    Switch agents without changing how you work. The conversation and control layer do not move when the underlying agent does.

  • 5
    Consistent workflow across projects and threads

    One workflow across conversations, projects, and agents — so switching context is a message, not a tab shuffle.

FAQ

Frequently asked

Can I run a coding agent while I sleep?

The underlying local coding agent performs the long-running work — NightMe keeps you connected and informed through the conversation. NightMe does not schedule overnight work or guarantee unattended completion; the agent’s own long-running behavior is what does that.

What does “babysitting” a coding agent mean?

Repeatedly checking, answering, redirecting, and planning work just to keep the agent-driven workflow moving. It is the loop of Observe → Respond → Direct → Continue that a developer falls into when the agent has no way to communicate state back.

How do I know when my coding agent needs me?

Through proactive IM message state, heartbeat / activity signals, and reaction emoji on the conversation itself. The statusbar on each bot message surfaces agent, tokens, git, and cwd — so you do not have to keep opening a terminal to check.

Do I need a separate app to monitor a coding agent?

No. NightMe uses supported familiar chat apps rather than requiring a dedicated NightMe communication app. The agent works on your machine; the conversation in your chat app is the monitor.

Can I use different coding agents for different tasks?

Yes — you can switch between any supported coding agent from the same conversation. NightMe is intentionally agent-agnostic and does not rank vendors; the point is that your workflow stays the same regardless of which agent is running.

Does NightMe run the coding agent itself?

No. The local coding agent remains responsible for execution — reasoning, tools, sessions, permissions, and whatever autonomous or long-running behavior it already supports. NightMe provides the conversation and control layer around it.

Your agents can change. Your workflow can stay the same.

Keep every project and session close at hand, and stay in the conversation wherever you are — no matter which coding agent is doing the work.

Sleep tight. NightMe codes all night.