Seu agent tem mãos agora (Drex Max)

Seu agent tem mãos agora (Drex Max)

What just happened (October 2026):

Nace AI released:

"Drex Max: Multimodal decision model for computer-use agents."

"Agents can now click, fill forms, automate workflows."

"Score: 67.59 (best in class)."

What this means for your business:

Before (current agents): Agent capability: Read text, call APIs Agent limitation: Can't interact with UIs

Workflow example: Customer: "Update my address in your system" Agent: "I'll update it for you" Agent (internally): Can call API to update (works)

BUT if customer has legacy system: Customer: "Update my address in legacy ERP" Agent: "I can't. I can't click on UIs" Customer: "Why not? It's just a form" Agent: "I'm limited to APIs" (frustration)

After (with Drex Max): Agent capability: Read text, call APIs, click UIs Agent capability: Fill forms, navigate browsers, automate workflows

Same workflow example: Customer: "Update my address in legacy ERP" Agent: "I'll do it" Agent (internally): 1. See login form (read text) 2. Type username (take action) 3. Type password (take action) 4. Click login (take action) 5. Navigate to settings (take action) 6. Fill address form (take action) 7. Click save (take action) 8. Return to customer Customer: "Done in 10 seconds!" (amazed)

Difference: Before: Agent helpless (can't use UIs) After: Agent omnipotent (can use ANY UI)

Impact:

• Automations possible: 10x more
• Workflows automated: Those requiring clicking
• RPA replacement: Agent can do what RPA bots do
• Agent value: Suddenly much higher

---

The problem Drex Max solves (UI gap)

Why current agents are limited (the gap)

Current agent capabilities:

What agents CAN do: ✓ Read text ("What's my order status?") ✓ Call APIs (Query database, update record) ✓ Process images (Read document, analyze) ✓ Generate text (Write response, explain) ✓ Reason (Decide what to do next)

What agents CANNOT do: ✗ Click on things ("Click the button") ✗ Fill forms manually ("Fill address field") ✗ Navigate UIs ("Go to settings page") ✗ Interact with legacy systems ("Use old ERP") ✗ Automate workflows requiring clicks

Why this matters: Problem: 30% of business workflows live in UIs (not APIs) Gap: Agents can't touch those workflows Impact: Companies can't automate half their work Result: RPA bots still required (expensive, brittle)

Real-world scenarios where agents fail

Scenario #1: Legacy system form-filling

Situation: Company has 20-year-old ERP system System has NO API (ancient technology) System has only UI (clicks, forms, navigation)

Current agent: Customer: "Process my invoice in the ERP" Agent: "I can't interact with the ERP UI" Result: Manual processing required (30 minutes)

With Drex Max: Customer: "Process my invoice in the ERP" Agent: 1. See login screen 2. Type credentials 3. Navigate to invoice module 4. Click "New invoice" 5. Fill form fields (vendor, amount, date) 6. Click submit 7. Confirm success Result: Automated (30 seconds)

Impact: Save 30 minutes × 100 invoices = 50 hours/month Value: R$5,000/month automated

Scenario #2: Multi-system workflow (no API bridges)

Situation: Workflow: Update customer in system A, then system B System A: Has API (modern) System B: Has no API, only UI (legacy)

Current agent: Update system A: ✓ Easy (API call) Update system B: ✗ Blocked (no API, can't click) Result: Manual step required for system B

With Drex Max: Update system A: ✓ Easy (API call) Update system B: ✓ Easy (agent clicks UI) Result: Fully automated workflow

Impact: 100% workflow automation (vs 50%)

Scenario #3: Portal navigation (multiple steps)

Situation: Customer needs: Download invoice PDF from vendor portal Portal: Requires login → navigate to invoices → find invoice → click download

Current agent: Customer: "Get my invoice from vendor portal" Agent: "I can't navigate portals (clicking)" Result: Manual download required

With Drex Max: Customer: "Get my invoice from vendor portal" Agent: 1. Open portal URL 2. See login page 3. Click email field 4. Type email 5. Click password field 6. Type password 7. Click login button 8. Wait for page load 9. See invoices list 10. Find invoice (by date, amount) 11. Click download 12. Return PDF to customer Result: Fully automated (20 seconds)

Impact: Zero manual work, instant delivery

Scenario #4: RPA-like automation (high-volume)

Situation: Company processes 100 expense reports/day System has no API for approval workflow Currently: Human clicks "Approve" on each one (2 hours/day)

Current agent: Process expense #1: ✗ Can't click approve Result: All 100 manual approvals needed

With Drex Max: Process expense #1: ✓ Agent clicks approve Process expense #2: ✓ Agent clicks approve ... Process expense #100: ✓ Agent clicks approve Result: All 100 automated

Impact: Eliminate 2 hours/day of clicking Value: 1 FTE freed up (R$5,000/month)

---

How Drex Max works (practical explanation)

What is Drex Max (definition)

Drex Max: A decision model that looks at screens

Traditional LLM: Input: "Click the save button" Output: Text ("I clicked the button") Problem: Can't actually click (just talks about it)

Drex Max (multimodal decision model): Input: Screenshot of UI + Task ("Click the save button") Process: 1. Read screenshot (see all elements) 2. Understand task (what needs doing) 3. Decide: What to click? Where? How? 4. Return decision (coordinates + action) Output: {action: "click", coordinates: [500, 200], confidence: 0.98} Result: Agent executes click

Advantage: Makes actual decisions about UIs (not guesses)

Key features:

1. Multimodal: Reads text AND images (UI screenshots)

   • Sees button labels
   • Sees form fields
   • Sees navigation menus
   • Understands layout

2. Calibrated probabilities: Returns confidence scores

   • "95% sure this is the save button"
   • "47% unsure about field name"
   • Agent can validate uncertain decisions

3. Typed output: Returns structured decisions

   • Not text ("click the button")
   • Actual decision (click coordinates, button ID)
   • Ready for execution

4. Efficient: One call per click

   • See screenshot → Decide click → Execute
   • No wasted tokens
   • Fast execution
   • Cheap (low token usage)

Drex Max vs alternatives

Comparison:

Option 1: Traditional LLM (GPT-4, Claude) Input: Screenshot + "Click save button" Output: "I see a button labeled Save. The button is blue." Problem: Describes UI (doesn't click it) Cost: Expensive (high token usage) Accuracy: ~60% (just guessing coordinates)

Option 2: Vision model (GPT-4V, Claude Vision) Input: Screenshot + "Click save button" Output: "Based on the screenshot, I would click at coordinates [450, 250]" Problem: Still generating text (not structured decision) Cost: Very expensive (multimodal token cost) Accuracy: ~70% (better than text, still guessing)

Option 3: Drex Max (specialized decision model) Input: Screenshot + "Click save button" Output: {action: "click", coords: [500, 200], confidence: 0.98} Advantage: Structured decision (ready to execute) Cost: Cheap (specialized, efficient tokens) Accuracy: 97%+ (trained specifically for UI decisions)

Winner: Drex Max (best accuracy, lowest cost, structured output)

---

Real-world implementation (how to use Drex Max)

Step 1: Choose your agent framework

Options:

1. Browser automation agents (recommended)

   • Framework: Playwright / Selenium
   • Agent: OpenClaw (built for this)
   • Model: Drex Max (decision-making)
   • Cost: R$50-500/month (depends on volume)
   • Timeline: Ready to use NOW

2. Custom agent (DIY)

   • Framework: Your own code
   • Agent: Your design
   • Model: Drex Max API (call directly)
   • Cost: API call pricing (~R$0.05 per decision)
   • Timeline: 2-3 weeks to build

3. RPA replacement

   • Replace: UiPath / Blue Prism
   • With: Drex Max agent
   • Advantage: Much cheaper, more flexible
   • Cost: 90% savings vs traditional RPA
   • Timeline: Migrate gradually

Step 2: Define workflows to automate

Workflow selection:

Best candidates (easy wins): ✓ Repetitive clicks (same workflow, many times) ✓ UI navigation (go to page A → B → C) ✓ Form filling (lots of forms, same fields) ✓ Multi-system workflows (copy data between systems) ✓ Legacy system automation (no API alternative)

Good candidates: ✓ Portal navigation (login, find, download) ✓ Report generation (click, wait, download) ✓ Data entry (type into forms)

Hard candidates (requires more logic): ✗ Complex decision trees (many if/else branches) ✗ Variable workflows (different paths each time) ✗ Real-time reaction (needs <100ms latency)

Not suitable: ✗ Machine vision tasks (image recognition, OCR) ✗ Real-time gaming (speed-critical) ✗ Complex reasoning (not a thinking model)

Step 3: Build agent workflow

Example: Invoice approval automation

javascript // Pseudo-code: Agent workflow using Drex Max

const approveInvoices = async (invoiceList) => { for (const invoice of invoiceList) { // Step 1: Open invoice portal const browser = await launchBrowser(); const page = await browser.newPage(); await page.goto('https://vendor-portal.com');

// Step 2: Agent sees login screen
let screenshot = await page.screenshot();
let decision = await drexMax.decide({
  screenshot: screenshot,
  task: "Login with user@company.com and password"
});
// Returns: {action: "click", coords: [200, 100], confidence: 0.96}

// Step 3: Execute decision
await page.click(decision.coords[0], decision.coords[1]);
await page.type(getEmailField(), 'user@company.com');

// Step 4: Agent sees password field
screenshot = await page.screenshot();
decision = await drexMax.decide({
  screenshot: screenshot,
  task: "Click password field and type password"
});
await page.click(decision.coords[0], decision.coords[1]);
await page.type(getPasswordField(), 'password123');

// Step 5: Agent sees login button
screenshot = await page.screenshot();
decision = await drexMax.decide({
  screenshot: screenshot,
  task: "Click login button"
});
await page.click(decision.coords[0], decision.coords[1]);
await page.waitForNavigation();

// Step 6: Agent sees invoice list
screenshot = await page.screenshot();
decision = await drexMax.decide({
  screenshot: screenshot,
  task: `Find invoice #${invoice.id} and click it`
});
await page.click(decision.coords[0], decision.coords[1]);

// Step 7: Agent sees approve button
screenshot = await page.screenshot();
decision = await drexMax.decide({
  screenshot: screenshot,
  task: "Click the Approve button"
});
await page.click(decision.coords[0], decision.coords[1]);

// Step 8: Verify approval
screenshot = await page.screenshot();
decision = await drexMax.decide({
  screenshot: screenshot,
  task: "Verify that invoice is now approved"
});
console.log(`Invoice #${invoice.id} approved: ${decision.confidence}`);

await browser.close();

} };

// Usage: await approveInvoices([ {id: "INV-001"}, {id: "INV-002"}, {id: "INV-003"}, // ... 100 invoices ]);

// Result: 100 invoices approved automatically (5 minutes) // Before (manual): 2 hours of clicking // After (Drex Max): 5 minutes, hands-free // Savings: 115 minutes per 100 invoices

Step 4: Monitor agent performance

What to track:

1. Success rate: % of workflows completed successfully Target: >95% Alert: If <90%

2. Confidence: Average confidence of Drex Max decisions Target: >0.90 Alert: If <0.80 (agent unsure)

3. Accuracy: Did agent click the right thing? Target: 99%+ Alert: If <95% (wrong decisions)

4. Speed: How fast does agent complete workflow? Target: <5 minutes per workflow Alert: If >10 minutes (too slow)

5. Cost: API calls × cost per call Target: <R$0.50 per workflow Alert: If >R$1.00 (expensive)

---

Cost-benefit analysis (why implement Drex Max)

Scenario: 100 invoices/day approval workflow

Current state (manual):

Task: Approve 100 invoices in legacy ERP system Method: Human clicks approve on each one Time per invoice: 1 minute (login already done) Time per 100: 100 minutes (1.67 hours) FTE cost: R$5,000/month Time cost per month: R$5,000 × 20 days × 1.67 hours = R$1,667 Annual cost: R$20,000 (just approval workflow)

With Drex Max agent:

Setup cost:

• Agent development: 1 week (R$5,000)
• Integration: 2 days (R$1,000)
• Testing: 2 days (R$1,000)
• Total setup: R$7,000

Operating cost:

• Drex Max API: R$0.50 per workflow
• Invoices: 100/day × 20 days = 2,000/month
• Monthly API cost: 2,000 × R$0.50 = R$1,000
• Annual API cost: R$12,000

Total first year: R$7,000 + R$12,000 = R$19,000 Savings first year: R$20,000 - R$19,000 = R$1,000

BUT:

• Freed up 1.67 hours/day per person
• 1 FTE can now do other work
• Value: R$5,000/month → R$60,000/year
• Net benefit: R$60,000 - R$19,000 = R$41,000/year

Final calculation: Year 1: -R$7,000 setup + R$41,000 savings = R$34,000 profit Year 2+: R$41,000 profit (no setup cost) ROI: 434% in year 1, infinite in year 2+ Payback: <3 months

Hidden benefits

✓ Error reduction: Manual approval = 2% error rate. Agent = 0.1% ✓ Compliance: Full audit trail (agent logged every action) ✓ Speed: 100 invoices in 5 minutes vs 100 minutes (20x faster) ✓ Scalability: Same agent can handle 1,000 invoices/day (just add more VMs) ✓ 24/7 operation: Agent works nights/weekends (no extra cost) ✓ No training: New employees don't need training on approval workflow

---

Drex Max vs alternatives (comparison)

Option 1: Traditional RPA (UiPath, Blue Prism)

Pros:

• Mature (proven, stable)
• Enterprise support

Cons:

• Expensive (R$2,000-5,000/month per bot)
• Brittle (breaks with UI changes)
• Slow to adapt (need developer for changes)
• Not intelligent (just replaying clicks)
• Licensing hell (per-process, per-machine)

Option 2: Manual labor (hire people)

Pros:

• Flexible (humans adapt to anything)
• Smart (humans think about edge cases)

Cons:

• Expensive (R$5,000+/month per person)
• Slow (40-60 hours/week of work)
• Error-prone (humans make mistakes)
• No 24/7 (no night shifts without extra cost)
• Training required (onboarding time)

Option 3: Drex Max agent (AI-powered)

Pros:

• Cheap (R$100-500/month for agent + API calls)
• Intelligent (understands UI, makes decisions)
• Fast (10x faster than manual)
• Accurate (99%+ when configured well)
• Scalable (same cost for 10 or 1,000 workflows)
• 24/7 (no extra cost for nights/weekends)
• Adaptable (can handle UI changes with retraining)

Cons:

• Newer technology (less proven than RPA)
• Still needs monitoring (not 100% hands-off)
• Complex UIs (might struggle with very complex interactions)

Winner: Drex Max

Cost comparison (per month, high-volume): RPA bot: R$3,000 Manual (1 FTE): R$5,000 Drex Max agent: R$300 Savings: 90-94% vs alternatives

---

When to use Drex Max (decision tree)

Is your workflow suitable for Drex Max?

Question 1: Does workflow involve clicking/typing in UIs? YES → Continue NO → Use APIs instead

Question 2: Is workflow repetitive (same steps each time)? YES → Perfect for Drex Max NO → Might be too complex

Question 3: Are there 10+ workflows per month? YES → High ROI NO → Might not be worth it

Question 4: Does workflow currently take >5 minutes? YES → High savings NO → Marginal savings

Question 5: Is there a legacy system (no API)? YES → Drex Max is only option NO → Consider APIs first

If you answered YES to 3+: → Implement Drex Max NOW (high ROI, fast payback)

If you answered YES to 2: → Implement Drex Max (good ROI, worth trying)

If you answered YES to 1: → Implement Drex Max (might be worth it)

---

Conclusion: Drex Max = RPA finally gets smart

Summary:

Drex Max changes the equation:

Before: ✗ Agents: Can't click (no UI interaction) ✗ RPA: Too expensive, too brittle ✗ Manual: Too expensive, too slow Result: 30% of workflows stay manual

After (with Drex Max): ✓ Agents: CAN click (full UI automation) ✓ Cost: 90% cheaper than RPA ✓ Speed: 20x faster than manual ✓ Quality: 99%+ accuracy Result: 100% of workflows can be automated

The question for your company:

Your workflows:

• Include clicking? (Probably yes)
• Include form-filling? (Probably yes)
• Involve legacy systems? (Possibly yes)
• Currently manual? (Definitely yes)
• Cost R$5,000+/month? (Probably yes)

Drex Max:

• Cost: R$300-500/month
• Savings: R$4,500+/month
• ROI: 10-15x
• Payback: 2-4 weeks

Will you: Option A: Keep doing it manually (waste money) Option B: Buy RPA bot (expensive, brittle) Option C: Use Drex Max agent (smart, cheap, fast)

→ OpenClaw: Drex Max integration, computer-use agents, browser automation agents, UI automation, workflow automation, legacy system automation, RPA replacement, agent framework

Seu agent: Finalmente tem mãos. Clica, preenche, automatiza. 🤖✋