DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///
DARC.DEV
All guides
WorkflowLevel · Start#Automation#n8n#Approval

An AI workflow that drafts, validates and waits for approval

Build a safer automation pattern for incoming requests: deterministic routing around a narrowly scoped model step, with idempotent writes and a visible error queue.

DCDariusz Czechowicz23 Sept 20263 min readPractical guide
24
Workflow
Source-checked Sources & corrections
4 practical sections · checklist · sources

Practical guide. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.

For: Builders automating document, support or content workflows

Before you start: A test integration account · A clear definition of an approved action

⚡ TL;DR — the 30-second version

  • Separate drafting from sending.
  • Approve exact content and recipient.
  • Treat retries and ambiguous delivery as first-class states.

A useful automation does not need an agent to decide everything. For an incoming support request, a model can propose a category and draft a response while ordinary workflow logic validates fields, checks permissions and asks for approval. This is a better first project than an agent that can read and send every email in your account.

The same pattern works in n8n, Dify, Flowise or hand-written code. The interface changes; the responsibilities do not.

“Put flexibility inside the model step and reliability around it.”
— Editorial interpretation

Field guide / 01

01 / Normalize and minimize the input

Assign a stable internal event ID before calling a model. Remove attachments or personal fields that are not needed for the task. Validate the input's size and type, and send unsupported formats to review instead of improvising a conversion.

For the first version, process only one known event type. Adding five connectors before you understand duplicate delivery makes debugging much harder.

Field guide / 02

02 / Ask for a proposal, not an action

The model should return a category, a draft and a review reason. It should not receive a send-email credential. Check the category against an allowlist and validate whether the evidence supports the draft.

Use an explicit unknown state for ambiguity. A workflow that forces every request into a category can confidently route a financial complaint to a generic support answer.

StageAllowed behaviorOn failure
IntakeNormalize and deduplicateError queue
ModelClassify and draftReview required
ValidatorCheck structure and policyReject proposal
ReviewerApprove exact contentEdit or decline
SenderExecute once with scoped credentialsRecord outcome, bounded retry

Field guide / 03

03 / Bind approval to the exact action

Show the recipient, message and relevant evidence to the reviewer. If the content changes after approval, require a new approval. An approval for 'respond to this customer' is broader and less auditable than approval for one exact message.

Timeouts and cancellations need defined behavior. An abandoned approval should not silently become permission after a retry or workflow resume.

Field guide / 04

04 / Replay the event before shipping

Submit the same event twice, interrupt after approval and simulate a provider timeout. Confirm that you create at most one external action and keep a record of uncertain delivery states. A timeout after sending may mean the action succeeded even if your application did not receive confirmation.

The deliverable is a flow diagram, three failure traces and a successful approval path. Only then expand to more event types or automatic approval for a narrow, tested class.

Before you ship

0/5

Use this as a review aid. Checks reset when you leave this page.

Put it into practiceBuild a bounded agent brief

Sources & verification

Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.

  1. 01n8n — integrating AI into workflowsn8n↗
  2. 02Dify — workflow quick startDify↗
  3. 03Flowise — assistants, chatflows and agentflowsFlowise↗
  4. 04MCP — tools, permissions and validationModel Context Protocol↗
Editorial policy and correction log →
#Automation#n8n#Approval#Agents