Refinement in loop
The same agent passes through the text three times, improving each round, and only then a second agent presents the result.
Updated on Sep 02, 2026
Asking "improve this text" once gives an improvement. Asking three times, each one on the previous version, gives something else — it's the same trick as someone rewriting a paragraph until it's good.
The While Loop card repeats a section of the flow while the condition is true, with a safety cap (max_iterations) so no agent runs forever.
How to import
On the platform: Agents → Import, choose the downloaded file. It enters as a new draft (nothing existing is changed), with refreshed observability IDs. Then fill in what belongs to your environment — credentials, URLs, and knowledge bases — and publish.
{
"format": "hinow.agent",
"version": 1,
"exported_at": "2026-09-01T00:00:00Z",
"credentials_included": false,
"agent": {
"name": "Refinamento em laço",
"description": "O texto passa três vezes pelo mesmo agente, melhorando a cada rodada, e só depois é apresentado.",
"avatar": null,
"model": "hinow/hicode",
"system_prompt": null,
"config": {},
"tools": null,
"workflow": {
"nodes": [
{
"id": "start",
"type": "start",
"position": {
"x": 300,
"y": 0
},
"data": {
"label": "Início",
"variables": []
}
},
{
"id": "loop-1",
"type": "while",
"position": {
"x": 285,
"y": 150
},
"data": {
"name": "3 rodadas",
"loop_condition": "{{round}} <= 3",
"max_iterations": 4
}
},
{
"id": "agent-melhora",
"type": "agent",
"position": {
"x": 560,
"y": 150
},
"data": {
"name": "lapidador",
"model": "hinow/hicode",
"system_prompt": "Melhore o texto mais recente da conversa (o do usuário na 1ª vez; depois, a sua última versão): mais claro, mais forte, mais conciso. Responda SÓ com a nova versão, prefixada por 'v{{round}}: '.",
"config": {}
}
},
{
"id": "agent-final",
"type": "agent",
"position": {
"x": 275,
"y": 340
},
"data": {
"name": "apresentador",
"model": "hinow/hicode",
"system_prompt": "Apresente a versão final do texto lapidado e, em 1 frase, o que evoluiu desde o original.",
"config": {}
}
},
{
"id": "end-1",
"type": "end",
"position": {
"x": 305,
"y": 500
},
"data": {
"label": "Fim"
}
}
],
"edges": [
{
"id": "e-start-loop-1-d",
"source": "start",
"target": "loop-1"
},
{
"id": "e-loop-1-agent-melhora-loop",
"source": "loop-1",
"target": "agent-melhora",
"sourceHandle": "loop"
},
{
"id": "e-agent-melhora-loop-1-d",
"source": "agent-melhora",
"target": "loop-1"
},
{
"id": "e-loop-1-agent-final-exit",
"source": "loop-1",
"target": "agent-final",
"sourceHandle": "exit"
},
{
"id": "e-agent-final-end-1-d",
"source": "agent-final",
"target": "end-1"
}
]
}
}
}| Card | Why it's here | What it does |
|---|---|---|
| **While Loop** | Controlled repetition. | Condition {{round}} <= 3 and max_iterations: 4. Two exits: LOOP (repeat) and EXIT (continue). |
**Agent polisher** | The work that repeats. | Improves the last version and prefixes with v{{round}} — you can see the evolution round by round. |
| **The loop back** | It's what closes the cycle. | The agent's connection **back** to the While is what makes the next round happen. Without it, the loop runs only once. |
**Agent presenter** | Separate iterating from delivering. | Exits via EXIT and presents the final version, saying in one sentence what evolved. |
Always with a cap
max_iterations is the safety belt: if the condition never becomes false (variable that doesn't increment, poorly written condition), it's what stops the loop. A loop without a cap is the fastest way to burn tokens.
- Until it passes the criterion, not until count 3: swap the condition for a variable that an evaluator agent sets (
{{approved}} == false) — the loop stops when quality arrives, not when the counter runs out. - Reprocess a queue: the loop consumes a collection declared at Start, one item per round.
- Retry with backoff: a loop around a Webhook that sometimes fails, with the agent deciding whether to try again.