How always-on
agents work

The week-one loopS03

messages = [system_prompt, user_message]
while True:
    reply = llm(messages, tools=TOOLS)
    messages.append(reply)
    if not reply.tool_calls:
        print(reply.text)
        break
    for call in reply.tool_calls:
        messages.append(run_tool(call))

• Their implementation: Bot (S08). Bots left (Sales Outbound, Inbox Manager...), chat right; lacking a Salesforce login, it asks you to sign in on its computer.

Six capabilitiesS10

1. Work for hours, following instructions throughout.
2. Hand off to helpers, each with a fresh context.
3. Use real tools: shell, files, browser, MCP.
4. Use a computer, even apps with no API.
5. Keep knowledge in files.
6. Learn by watching: show it once, it saves a skill.

• Cloud computer (S15): one Firecracker VM per user; Debian container kept alive by a supervisor. Files, installed tools, browser logins persist across sessions.
• Why the loop runs on the server (S17): laptop closed, work continues; seamless computer–phone switching; no App Store review cycles; months-long agents can't live on devices^◐.
• Key point (S18): state lives on the server, so the VM can be rebuilt anytime (e.g. for a Linux 0-day) without stopping the loop.

What happens when you hit sendS19

1. Sending: shown locally at once with a unique ID; the server drops repeated IDs, so retries are safe.
2. The turn: joins the queue; one turn at a time per conversation; every step saved to the database.
3. Watching: the server sends only a "something changed" signal; apps fetch the data, so a lost signal is harmless.
4. Ordering: every update has a sequence number (41, 42, 43); gaps reload from the database^◐.

Three ways to use a computerS27

• Read the page structure: list buttons and links; fast and reliable.
• Click on pixels: a 1280×800 screen, for desktop apps and system dialogs.
• Ask a person: hand over at logins, two-factor codes, captchas, payments.
• A helper does the clicking, keeping screenshots out of the main context.

4 tips for tool designS29

1. Don't add or remove tools between turns.
2. Keep each tool's description and schema under 4,000 characters.
3. Save state with a dedicated tool, so it can be reviewed.
4. Errors say what to do next: "The computer is asleep. Use Shell or Read to wake it, then try again."

• Main context keeps only pointers or summaries (S33): long outputs to files, screenshots and heavy work to helpers, old messages to summaries, instructions to skills.
• Treat a cache miss like a bug (S34): prefix unchanged until the next summary; deploys send running conversations a note, not new tool text; fixed order; per-section hashes; cache warmed as you type.

• Three training stages (S37):
  • Pre-training: language, code, general knowledge.
  • SFT: learns the harness: tools and formats, when to reply quickly.
  • RL: practices long instructions, skills and memories, clicking the right spot (grounding), working with other bots.
• Inference is tuned separately: the main agent replies in seconds; helpers reliably run for hours^◐.

• Architecture compounds: when code is cheap, the right architecture keeps you faster than everyone else.
• Build your own: start from the week-one agent; add a server, a computer, memory. Open-source always-on agents are great to learn from.
• Learn LLM fundamentals: like knowing databases when building a website^◐.