The promise of AI-powered web agents automating tedious browser workflows has long been undermined by a harsh reality: dazzling demos crumble in production. As Calvin French-Owen observed, tools like GitHub Copilot succeed because they avoid flakiness—a pitfall plaguing web agents. Simplex’s early attempts saw <10% success rates for basic tasks like downloading batches of checks from portals. The culprit? Not prompt engineering, but context engineering—the art of structuring all inputs for solvability.

Why Web Agents Drown in Context

Traditional web agents operate with ballooning context comprising:
1. User Task: The original goal (e.g., “Download checks”)
2. Web Page Content: A text snapshot of the current state (up to 30K tokens!)
3. Agent Memory: Accumulating logs of past actions/results

This architecture triggers two critical failures:

| Problem                | Consequence                          | Impact                          |
|------------------------|--------------------------------------|---------------------------------|
| Memory Accumulation   | Linearly growing context (11K+ tokens) | "Context confusion"—past errors haunt current decisions |
| Page Content Domination | 88%+ context consumed by page state  | Agent loses sight of core task  |

In workflows exceeding 10 steps, agents fixate on outdated modal errors or drown in dropdown options. Reliability plummets.

The Multi-Agent Breakthrough

Simplex’s solution draws from Anthropic’s research: a lead orchestrator agent spawning focused sub-agents.


alt="Article illustration 1"
loading="lazy">

**How it works**: 1. The lead agent holds the **long-term goal** (e.g., “Process 50 invoices”). 2. It spawns short-lived sub-agents for micro-tasks: - `Extract invoice list` - `Download checks for Invoice X` - `Navigate between views` 3. Sub-agents execute tasks, then return **compressed results** (e.g., "Invoice 1234: 5 checks downloaded") before terminating. 4. The lead agent retains only summarized context, not raw page data. 
<img src="https://news.lavx.hu/api/uploads/taming-context-chaos-how-multi-agent-architectures-solve-web-automations-reliability-crisis_20250730_154040_image.jpg" 
     alt="Article illustration 2" 
     loading="lazy">

Results: From 10% to Hour-Long Reliability

Metrics reveal drastic improvements:
- Context Tokens: Lead agent memory stabilized at ~4K tokens vs. uncontrolled growth
- Success Rate: 50+ invoice workflows completed flawlessly vs. failing at 5-10 invoices previously
- Duration: 60+ minutes of continuous operation (Video demo)

By isolating context—sub-agents handle dense page states, the lead agent focuses on strategy—Simplex bypasses the pitfalls of monolithic architectures. This isn’t just theory; it’s enabling enterprises to automate revenue-critical workflows like financial document processing.

Beyond the Demo

While multi-agent design solves ~70% of reliability issues, the battle continues. Browser quirks, evaluation frameworks, and adversarial page structures remain challenges. Yet this architectural shift proves web agents can cross the chasm—when we stop treating them as single LLMs and start engineering context like distributed systems.

Source: Simplex Blog - Context Engineering for Web Agents