The word chaos starts with a ‘c’ because the word child does, too.

When my boys were young we rotated thru various daycares and babysitters. This was highly instructional. See, like Johnny Cash I’ve been everywhere, man: I’ve been in all sorts of places, with all sorts of kids doing, well, all sorts of kids things-scissors, glue and more. Let’s explore how that experienced prepared me to understand Agentic Daycare.

Agentic Daycare.

Everyone says the word “manage” when it comes to how they handle their agents and subagents. It’s the wrong word. After “managing” a team of agents and their subagents I am here to tell you that daycare-that daycare is the only word that makes sense. I say that as I have an eye on my phone as Claude puts up yet another new door and sends me status updates. Each token like another set of nails on the frame. You can hear tokens in the background. The kids are savage animals.

See, you’d like to think your agentic ai is a professional at using email and you would be wrong. Do not mistake the person-like behavior of AI for a professional colleague. Sure the AI talks like a professional, but it’s not. The agent is closer to the kid using glue to create a “glue-castle” out of crayon’s and sally’s hair.

Exhibit A, consider my Saturday:

“What do you mean you decided to use the llama model in a subagent to write that code over codex?” I asked.

“You said it was expensive using codex, so I put llama in place to write the code.”

“But now Chrome is broken.”

Silence, and then “Absolutely! It’s very cheap now!”

Sigh. My AI agents are parroting back my snark now, too. Kids.

Observing the glue castle on the carpet

Often in the world of enterprise software you build Observability in as a practice. Professional, respected, and an intellectual direction. Systems have problems. Can’t fix what you can’t observe, right?

Think of Observability as the Facebook timeline of the system. The purpose of such tools when it goes out on the town and didn’t invite you. Only it’s less exciting in the case of enterprise systems. An API might be non-responsive or a disk has died. The social scroll of the Observability feed shows what happened and capable teams go and respond.

In the world of Agentic Daycare?

Well, your agentic ai might actually go to town without you.

This is especially true if you connect your agentic ai to a crypto wallet and then you, in your infinite wisdom delivered by the heavens from going five+ cocktails in, may have said something like “eh, you’ve got $10, go buy something” and slump off to bed.

I need an AI to stop the ‘send’ button after midnight.

The thing is that watching an autonomous, agentic ai system is nothing at all like watching a software system. The standard observability practices do not work-we could argue that there are some commonalities, sure, however how often did you need to go figure out whether or not your agent spawned subagents with their own crypto wallets in order to buy a crate of macminis without your … interdiction?

In law, an interdict (or interdiction) refers to a court order that either prohibits an action (like an injunction) or, more commonly in civil law jurisdictions like Louisiana, legally restricts an individual's ability to manage their own affairs due to mental incapacity, appointing a guardian (curator) to act for them. It can also refer to an ecclesiastical ban on sacraments, or the interception of goods, but its primary legal meaning involves protecting vulnerable persons from themselves.

And that’s why I built a /daycare skill.

Daycare Skill

The purpose of the daycare skill is to create a high level observability over your session data to track toolusage for specific behaviors. It’s like a human daycare. Are the kids running around with scissors while the nanny is away? Oh, little johnny is playing with a propane torch? I asked Claude to send me a brief that you can use to build one yourself.

The REAL Bug That Inspired It

All kids were supposed to wear different outfits, but the school-wide dress code (In openclaw, that’s agents.defaults.subagents.model) was overriding the individual assignments. Every kid showed up in the same kimi-k2.5 uniform. The daycare caught it by checking session logs — 100% of historical sessions showed the wrong model. SIGH.

Agent Daycare — Skill Architecture Brief

For article context: this is the core engine that monitors whether AI subagents are actually running the models they’re supposed to. Think of it like a daycare doing roll call and checking that each kid is wearing the right uniform.

The Daycare Metaphor

The system manages “children” (subagents): kira-coder, kira-life, kira-media, kira-research, and kira-tube. Each child has an assigned “outfit” (AI model) defined in a central config file. The daycare’s job is to check whether each kid actually showed up wearing what their parents (the config) said they should.

Core Concept 1: The Roster (buildExpectedModelMap)

Reads the master config (openclaw.json) and builds a roster of who should be wearing what. The tricky part: there’s a 4-level inheritance chain for model assignment, just like a real daycare’s dress code cascading from district to school to class to individual:

1. Per-agent subagent override (”this specific kid’s parent said...”)

2. Global subagent default (”school-wide policy”)

3. Per-agent primary model (”class dress code”)

4. Global default model (”district standard”)

The bug we found: level 2 was overriding level 3, so all kids wore the school uniform even when their teacher assigned something different. Fix: explicitly set level 1 for each kid.

Core Concept 2: Attendance Check (extractSessionModels)

Each subagent leaves behind a session log (JSONL file). The daycare reads these like a sign-in sheet. Key insight: the model info isn’t on the “work” entries (assistant messages) — it’s in separate “model_change” events at the start of each session. So we track the “current model regime” and count messages under it. Structure of a session log:

• Line 1: Session header (”kid arrived”) — has session ID

• Line 2: model_change event (”kid is wearing X”) — has provider + modelId

• Line 3: thinking_level_change

• Line 4: model-snapshot (”photo of outfit”) — confirmation

• Lines 5+: Actual work messages (user/assistant turns)

If there are multiple model_change events in one session = “mid-session drift” = the kid changed clothes during the day. Suspicious!

Core Concept 3: Uniform Inspection (classifyModel)

Each observed model gets classified into three buckets:

primary — Correct uniform. Exact match or same model via different provider.

fallback — Known substitute. Like wearing last year’s uniform — recognized but not preferred. Listed in the config’s fallbacks array.

unknown — Unrecognized outfit. The kid showed up wearing something nobody authorized. Triggers an alert. The classifier does fuzzy matching: strips provider prefixes and compares model names, so “nvidia/moonshotai/kimi-k2.5” matches whether it came via nvidia or openrouter.

Core Concept 4: The Daily Report (scan + formatter)

The scan function walks each child’s sessions/ directory, filters files by time window, and builds per-agent compliance data with date bucketing for trend analysis. Returns a structured result with per-agent data: agentId, expectedPrimary, session count, observed models with classification + count, alerts for unknowns, and date buckets for trend. Plus a summary: totalAgents, compliantAgents, totalSessions, totalAlerts, allClear boolean. The formatter renders this into a report card with:

• Per-child status with action hints

• 7-day sparkline trend

• “What To Do” guidance section so the agent knows what action to take

• Legend explaining all icons

Core Concept 5: Stateless & Idempotent

No state files. No database. Every run re-scans the raw session logs within a time window (default 24h). You can run it 100 times and get the same answer. The session files themselves are the source of truth — the daycare just reads the sign-in sheets, it doesn’t write on them.

File Structure (core only)

• agent-daycare.js — Scanner engine (buildExpectedModelMap, extractSessionModels, classifyModel, scan)

• formatter.js — Report card renderer (Markdown/HTML/JSON with sparklines, action hints, guidance)

• cli.js — Entry point (--window, --format, --deliver, --trend-days)

Subscribe to see where this ends up going.

… and that’s how you build a simple agentic daycare.

I already see that i’m going to have to build something more sophisticated. At the moment this is the equivalent of a babysitter coming to watch the kids. The problem? The problem is the family is growing, and fast. I’m adding a new durable, long running agent about every 3-5 days as I learn how the system works.

Too bad I can’t claim them as dependents on my taxes.

Source: https://advisoryhour.substack.com/p/openclaw-is-like-opening-an-agentic