Letting Go of the Handlebars: Agents
By the end you'll be able to
- Explain an agent as a model running a repeated loop — plan, act via tools, observe results, revise — toward a stated goal.
- Explain why small per-step error rates compound across many autonomous steps.
- Decide where human checkpoints belong in a delegated task.
- Predict
- Gate
- Decide
- Field note
⚡ 30-second warm-up — From Lesson 14: what does the model actually do when it "uses a tool"?
It writes a structured work order; separate software executes it; the result comes back onto the desk. Today, that single call becomes a repeated loop — the model calling itself back into the conversation, again and again.
"Find me three quotes for repairing my fence, compare them, and draft replies." That's twenty steps, some in the world. Predict the first place it goes wrong.
Start with a delegated afternoon
You've used AI that answers a question. Now imagine handing it a whole task with many steps, some of which touch the real world — and walking away. What could possibly go wrong, twenty steps in?
Commit to a prediction
If a 20-step delegated task goes wrong, the error most likely first appears…
This lesson changes one thing: you'll stop picturing autonomy as a new kind of intelligence, and start seeing it as the pipeline you already know, wrapped in a loop — which means its strengths and its one big weakness are both entirely predictable.
Discover the mechanism
An agent is everything from this act composed into one architecture. Give the system a goal and permission to loop: it drafts a plan (generated text, Lesson 9), executes step one as a tool call (Lesson 14), reads the result back onto the desk (Lesson 10), and generates the next step in light of it — around and around, sometimes retrieving documents (Lesson 13), sometimes writing notes to itself (Lesson 15), until it judges the goal met.
Nothing new was added; the loop is the novelty. Its strength is tireless multi-step work at scale. Its structural weakness: each cycle treats the previous cycle's output as trustworthy context — so a plausible-but-wrong step two becomes the foundation of steps three through twenty (Lesson 4's raffle, compounding). This is why agent design is mostly about checkpoints, permissions and reversibility.
real systems vary in how plans are represented and revised, and multi-agent setups add further complexity — this lesson's Go deeper opens both. Lesson 19 completes the picture: the same scaffolding that gives an agent its reach is also its attack surface.Impressive on well-defined projects; but every misunderstanding they don't check with you gets built upon — by 4pm, a small early misreading has become thirty wrong phone calls. Good managers of temps define the task tightly, require check-ins at the risky points, and review before anything irreversible.
- The project brief = the goal prompt. Phone = tools (Lesson 14). Notes accumulating = the desk growing each cycle.
- Check-ins = human approval gates. The compounding misunderstanding = per-step errors feeding forward as trusted context.
- Where it breaks: a temp feels doubt and escalates; the model's plausibility machinery (Lesson 1) generates confidence by default, so agents typically barrel forward unless engineered to stop. A temp also learns your workplace permanently; the agent's competence resets every time (Lesson 5), keeping only what's on the desk or in notebooks (Lesson 15).
Run mission control yourself
Conceptual illustration — a scripted fence-quote mission, deterministic. Choose where to place approval gates before launching, then run it and watch the trace.
Where this bites at work
Luka trials an agent that monitors grant portals, retrieves matching rounds and drafts summaries — with a rule that nothing is submitted or sent without sign-off, and a step-and-spend cap: every cycle runs the full metered pipeline (Lesson 9), so an agent that loops is a bill that loops too.
Delegate the loop, keep the trigger, cap the meter.
"Agents are a new kind of autonomous intelligence" is the misconception to retire. They are the same prediction machine composed with tools in a loop; the autonomy is an engineering permission, adjustable and revocable — and choosing its level is a human decision.
Now make the call
Luka's grant-monitoring agent flags a promising match and drafts a full application summary, citing the funder's portal page. What does he do?
Try trusting the summary purely because it cited a source too — worth seeing that a citation and a verified claim are two different things.
Act 3 challenge: the "zero supervision" pitch
Your CEO forwards a vendor email: "Our autonomous agent handles your entire customer inbox — reads messages, retrieves policy answers, books callbacks, and replies — with zero supervision needed." — "Worth a trial?" Everything from this act, one decision at a time.
Apply it to your work
Find one multi-step AI-assisted task at your work and write down: where's the checkpoint before anything irreversible happens, and what's the cap on how many steps or how much spend it can run before a human looks at the trace?
Prove it to yourself
Two quick questions — untimed, retryable, and they count toward ★ Mastered (completed + decision + self-check all correct). This is what defeats the feeling of "that made sense" evaporating by next week.
Explain it in your own words
Two or three sentences for a colleague. Then compare against the rubric — it checks ideas, not wording, and only you see your note.
✓ Lesson 16 stamped: Loop Commander — Act 3 complete
You designed the oversight, ran the mission, and watched the difference gate placement makes. You now command the whole scaffolding architecture. Act 4 asks the question that keeps professionals employed: when this beautiful machinery fails, why — and how do you see it coming?
Evidence & review — how we know what this lesson claims
| Claim | Type | Basis | Review risk |
|---|---|---|---|
| An agent is a model executing a repeated plan–act–observe–revise loop with tool access | How it works | Architecture / primary research literature | Low |
| Per-step error rates compound across autonomous steps because each cycle treats prior output as trustworthy context | How it works | Primary research literature | Medium |
| Agent capabilities, protocols and safety practices are a fast-moving area | Product behaviour | Provider / industry documentation | High — reviewed quarterly |
| The mission-control demo | Teaching device | Deterministic, scripted illustration | Medium |
| The office-temp analogy | Useful mental model | Limits stated in the lesson | Low |
Concept review due: January 2027. Product-behaviour claims: quarterly.
Where you are in the machine
Act 3 complete — the whole scaffolding, understood. Select any layer for its role, lessons and your live progress.
The machine, bottom to top: training data (2) → token pieces (3) → prediction engine (1, 4–9) → the desk (10–11) → senses (12) → live knowledge (13) → tools (14) → memory (15) → agent loops (16) → human judgement (17–21).
Next: why this fails, and how to see it coming
You now command the whole architecture. Act 4 asks the question that keeps professionals employed: when this beautiful machinery fails, why — and how do you see it coming? Starting with the machine's most convincing failure: sounding certain while being wrong.