The Visualizer: TripDesk in action
TripDesk is a travel reservation agent . Go step by step and observe each beat of the loop: Perceive To reflect on Act Observe. You will also see a human checkpoint, a breakdown… and the shutdown decision made by the agent itself.
🧰 Tools:
chercher_vols · consulter_agenda · reserver_vol (irreversible!) envoyer_confirmation
The execution log will appear here, step by step…
🗒️ Agent notepad (working memory)
- Empty — execution has not started.
💡 Three things to notice
1️⃣ The counters top: max iterations and budget cap — two guardrails that prevent a loop from running infinitely.
2️⃣ The notepad (scratchpad) which is enriched: it is the working memory — the history reread at each iteration. She is erased at the end of the execution.
3️⃣ The END : no one stops TripDesk — it itself assesses that the goal has been achieved. This is THE difference with a chatbot.
Do you really need an agent?
Agents are powerful… and expensive: each iteration = one call to the model = cost + latency + risk. Answer the questions to find the simplest architecture that is enough . Golden rule: prompt < unique tool < fixed workflow < agent .
The essence of the session
The definition in one sentence
An agent = a language model in a loop, with tools, and the right to decide when it's finished.
It is not a “smarter” model: it is often exactly the same model than a chatbot, placed in a different architecture. Agency is a property of system .
Chatbot vs agent
| Chatbot | Agent | |
|---|---|---|
| Functioning | One turn: question → answer → stop | A loop, until the end of the task |
| Stop | Mechanics, after the answer | Decided by agent when the goal is achieved |
| Production | Text | Actions (tools) + text |
| Error | Wrong answer | Failed action → possible recovery |
The loop and its bricks
| Brick | Role | TripDesk example |
|---|---|---|
| 🔁 Loop | Perceive → reflect → act → observe, in cycle | 5 iterations from mission to confirmation |
| 🛠️ Tools | The agent’s “hands” = the “act” step (tool calling, Session 5) | chercher_vols , reserver_vol … |
| 🗺️ Planning | Break down the goal into sub-steps — plan revisable | Plan changes when agenda conflict is discovered |
| 🗒️ Working memory | History/notepad reread at each iteration — deleted at the end | The 3 flights found in iteration 1 remain available in iteration 4 |
| 🚧 Guardrails | Max iterations · budget ceiling ⚠ · human validation of irreversible actions | Max 15 iterations, €0.50 ⚠, validation before reserver_vol |
| 🩹 Error recovery | Read error → adapt → try again differently | “Full flight” → choose the alternative, not the same flight |
⚠ The amounts and numbers of iterations are orders of magnitude: the prices of APIs (Application Programming Interface) and practices evolve quickly.
When NOT to use an agent
If a single well-written prompt is enough (summary, translation), or if a single tool call is enough (weather), or if the steps are fixed and known in advance (workflow), an agent is there. over-engineering : more expensive, slower, riskier, for zero profit.
“Always take the simplest weapon that accomplishes the mission. »