Personal & Consumer Use Cases Basics

8 concept-first examples for a tiny personal agent - 5 basic and 3 intermediate.

You will model inbox triage, a chat-channel reply, approval for sends, and hard stop conditions. Code stays sketch-level so the use-case shape stays clear.

Prerequisites

• Comfort with the agent loop idea (model decides, tools run, observations return)
• Optional: Python 3.11+ if you want to type the sketches locally
• Treat mail and chat credentials as production secrets even in demos

# Optional local env for later experiments (not required to learn the patterns)
python -m venv .venv && source .venv/bin/activate
pip install pydantic

Basic Examples

1. Describe the Goal, Not the Product

Write a personal-agent goal as a checkable outcome.

Goal: Post a morning digest of (a) unread mail from VIP list and
(b) calendar events in the next 24 hours to my Telegram chat.
Stop when digest is posted or after 6 tool turns.

• Success is observable without "vibes."
• Stop conditions travel with the goal.
• No vendor name is required to scope the work.

Related: What a Personal AI Agent Actually Does Day to Day

2. Split Read Tools from Write Tools

Register tools with different risk levels.

from pydantic import BaseModel, Field
 
class ListUnreadArgs(BaseModel):
    max_messages: int = Field(default=20, ge=1, le=50)
 
class DraftReplyArgs(BaseModel):
    thread_id: str
    body: str
 
class SendReplyArgs(BaseModel):
    thread_id: str
    body: str
    approved: bool = False
 
# Read tools: list_unread, get_calendar_day
# Write tools: send_reply, create_event - only when approved is True

• The model can draft freely if send is blocked without approved.
• Descriptions should say when to use each tool.
• Day-one agents often ship read-only.

3. Simulate a Morning Mail Check

Walk one turn sequence on paper or in logs.

User/schedule: "Morning digest"
→ list_unread(max_messages=20)
← [{from: boss, subject: "Budget"}, {from: newsletter, ...}]
→ get_calendar_day(date=today)
← [{10:00 Standup}, {15:00 Dentist}]
→ final: "VIP: Budget email. Today: Standup 10:00, Dentist 15:00."

• Observations drive what goes in the digest.
• Newsletters can be filtered by rules or model judgment with a VIP allowlist.
• No send tools were needed for value.

4. Post the Result to a Chat Channel

Treat chat as a delivery tool, not as the whole agent.

def post_chat(channel: str, text: str) -> dict:
    # Platform SDK or webhook; return message id + ok flag
    return {"ok": True, "channel": channel, "chars": len(text)}
 
# Observation for the model:
# {"ok": true, "channel": "telegram:personal", "chars": 180}

• Chat integration is I/O.
• Keep messages short; link out if details are long.
• Rate-limit proactive posts so the agent does not spam you.

Related: Chat-Integrated Agents

5. Bound Every Wake with max_turns

Personal agents wake often. Each wake needs a budget.

def run_personal_wake(goal: str, tools: dict, max_turns: int = 6) -> str:
    state = {"goal": goal, "trace": []}
    for turn in range(max_turns):
        decision = decide(state, tools)  # model call
        if decision.done:
            return decision.answer
        obs = tools[decision.tool](**decision.args)
        state["trace"].append({"turn": turn, "tool": decision.tool, "obs": obs})
    return "stopped: max turns - partial digest not posted"

• Six turns is a reasonable first default for digests.
• Controlled stop beats silent hang.
• Log the trace; it is how you debug wrong tool choice.

Intermediate Examples

6. Draft a Reply, Wait for Approval

Never auto-send on day one.

User: "Reply to budget email: I'll send numbers by Friday."
→ get_thread(thread_id=...)
← {body, participants}
→ draft_reply(...)  # stores draft only
← {draft_id, preview}
→ final: "Draft ready. Reply APPROVE <draft_id> to send."
# Later message: APPROVE d_123
→ send_reply(thread_id=..., body=..., approved=True)

• Human-in-the-loop is a feature for personal mail.
• Separate draft storage from send execution.
• Log who approved and when (you).

Related: Email and Calendar Agents

7. Handle Tool Auth Errors as Observations

Expired Google tokens are normal life, not edge cases.

def list_unread(max_messages: int = 20) -> dict:
    try:
        return {"messages": fetch_mail(max_messages)}
    except AuthError:
        return {
            "error": "auth_expired",
            "action_needed": "Reconnect mail OAuth in settings.",
        }

• The model should tell you to reconnect, not invent an empty inbox.
• Do not put refresh tokens into the prompt.
• Count auth errors separately in your own notes or metrics.

8. Prefer a Preference File Over "Remember Everything"

Store durable personal rules outside the raw chat log.

PREFS = {
    "timezone": "America/Los_Angeles",
    "quiet_hours": ["22:00", "07:00"],
    "vip_senders": ["boss@example.com"],
    "tone": "concise, no emojis",
}
 
# Inject a short prefs block into system context each wake.

• Preferences are product requirements for a personal agent.
• Full-history memory is optional and privacy-heavy.
• Update prefs deliberately; do not let the model rewrite them freely without review.

Related: Why Self-Hosting a Personal Agent Appeals to Privacy-Conscious Users

Related

• What a Personal AI Agent Actually Does Day to Day - daily pattern overview
• Chat-Integrated Agents - channel surfaces
• Email and Calendar Agents - triage and scheduling depth
• Shell-Access Personal Agents - when local power is too much
• Personal & Consumer Use Cases Best Practices - access checklist before real accounts
• Self-Hosted Personal Agents Basics - runtime-focused follow-on

Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana @solana/web3.js + @solana/spl-token.