-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathdesign.html
More file actions
484 lines (459 loc) · 27.8 KB
/
design.html
File metadata and controls
484 lines (459 loc) · 27.8 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Why Design Matters | MAFBench</title>
<script src="https://cdn.tailwindcss.com"></script>
<link rel="stylesheet" href="styles/main.css">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/css/all.min.css">
</head>
<body class="bg-white">
<!-- Navigation -->
<nav id="navbar" class="fixed top-0 w-full bg-white/95 backdrop-blur-sm shadow-sm z-50">
<div class="max-w-7xl mx-auto px-4 sm:px-6 lg:px-8">
<div class="flex justify-between items-center h-16">
<div class="flex items-center">
<a href="index.html" class="text-2xl font-bold text-indigo-600">MAFBench</a>
</div>
<div class="hidden md:flex space-x-8 items-center">
<a href="index.html" class="text-gray-600 hover:text-indigo-600">Home</a>
<a href="design.html" class="text-gray-900 font-medium hover:text-indigo-600">Why Design Matters</a>
<div class="dropdown relative">
<a href="architecture/index.html" class="text-gray-600 hover:text-indigo-600 flex items-center">
Architecture Guide
<i class="fas fa-chevron-down ml-1 text-xs"></i>
</a>
<div class="dropdown-menu">
<a href="architecture/index.html">Overview</a>
<a href="architecture/single-agent.html">Single-Agent Characteristics</a>
<a href="architecture/multi-agent.html">Multi-Agent Characteristics</a>
<a href="architecture/environment.html">Environment</a>
</div>
</div>
<a href="results.html" class="text-gray-600 hover:text-indigo-600">Results</a>
<a href="mafbench.html" class="text-gray-600 hover:text-indigo-600">MAFBench</a>
<a href="paper.html" class="text-gray-600 hover:text-indigo-600">Paper & Code</a>
<a href="about.html" class="text-gray-600 hover:text-indigo-600">About</a>
</div>
<button id="mobile-menu-btn" class="md:hidden text-gray-600">
<i class="fas fa-bars text-xl"></i>
</button>
</div>
</div>
<div id="mobile-menu" class="hidden md:hidden bg-white border-t">
<div class="px-4 py-4 space-y-3">
<a href="index.html" class="block text-gray-600">Home</a>
<a href="design.html" class="block text-gray-900 font-medium">Why Design Matters</a>
<div class="pl-2 border-l-2 border-gray-200">
<a href="architecture/index.html" class="block text-gray-900 font-medium mb-2">Architecture Guide</a>
<div class="pl-4 space-y-2">
<a href="architecture/index.html" class="block text-gray-600 text-sm">Overview</a>
<a href="architecture/single-agent.html" class="block text-gray-600 text-sm">Single-Agent Characteristics</a>
<a href="architecture/multi-agent.html" class="block text-gray-600 text-sm">Multi-Agent Characteristics</a>
<a href="architecture/environment.html" class="block text-gray-600 text-sm">Environment</a>
</div>
</div>
<a href="results.html" class="block text-gray-600">Results</a>
<a href="mafbench.html" class="block text-gray-600">MAFBench</a>
<a href="paper.html" class="block text-gray-600">Paper & Code</a>
<a href="about.html" class="block text-gray-600">About</a>
</div>
</div>
</nav>
<!-- Hero Section -->
<section class="pt-32 pb-16 px-4 sm:px-6 lg:px-8 bg-white border-b border-gray-200">
<div class="max-w-4xl mx-auto">
<h1 class="text-5xl md:text-6xl font-bold text-gray-900 mb-6">
Why Design Matters
</h1>
<p class="text-2xl text-gray-700 mb-6">
The hidden bottleneck isn't your model. It's your architecture.
</p>
<p class="text-lg text-gray-600 leading-relaxed">
You can have the best LLM, the perfect prompts, and unlimited context. But if your architecture is wrong,
you'll get 100× slower performance, 30% accuracy drops, and coordination failures. The good news?
Architecture is something you control. These results show exactly how to design systems that scale.
</p>
</div>
</section>
<!-- The Misconception -->
<section class="py-20 px-4 sm:px-6 lg:px-8 bg-gray-50">
<div class="max-w-6xl mx-auto">
<h2 class="text-4xl font-bold text-gray-900 mb-4">The Costly Misconception</h2>
<p class="text-xl text-gray-700 mb-12 leading-relaxed">
Most teams optimize the wrong things. Here's what actually moves the needle.
</p>
<div class="grid md:grid-cols-2 gap-16">
<div class="bg-white border border-gray-200 rounded-lg p-8">
<h3 class="text-2xl font-semibold text-gray-900 mb-6">What People Optimize</h3>
<ul class="space-y-3 text-gray-700 text-lg">
<li>• Models (marginal gains)</li>
<li>• Prompts (diminishing returns)</li>
<li>• Context length (expensive, limited impact)</li>
</ul>
</div>
<div class="bg-indigo-50 border-2 border-indigo-200 rounded-lg p-8">
<h3 class="text-2xl font-semibold text-gray-900 mb-6">What Actually Dominates</h3>
<ul class="space-y-3 text-gray-700 text-lg">
<li>• Orchestration structure (100× impact)</li>
<li>• Memory architecture (stability vs collapse)</li>
<li>• Planning interfaces (30% accuracy swings)</li>
<li>• Communication topology (90% → 30% success)</li>
</ul>
</div>
</div>
</div>
</section>
<!-- Controlled Architecture Evidence -->
<section class="py-20 px-4 sm:px-6 lg:px-8 bg-white">
<div class="max-w-6xl mx-auto">
<div class="bg-blue-50 border-l-4 border-blue-400 p-4 mb-8 rounded">
<p class="text-sm text-gray-700">
<i class="fas fa-info-circle text-blue-600 mr-2"></i>
<strong>Note:</strong> All numerical values and tables on this page should be verified against the paper PDF for the most accurate and up-to-date results.
</p>
</div>
<h2 class="text-4xl font-bold text-gray-900 mb-6">The Proof: Architecture Alone</h2>
<p class="text-lg text-gray-700 mb-8 leading-relaxed">
Same LLM. Same task. Only architecture changes. The results are stark: architectural choices alone
create order-of-magnitude performance differences. This isn't theory—it's measured across multiple
frameworks with controlled experiments.
</p>
<div class="bg-white border border-gray-300 rounded-lg overflow-hidden mb-6">
<table class="w-full text-sm">
<thead class="bg-gray-100">
<tr>
<th class="px-6 py-3 text-left font-semibold text-gray-900 border-b border-gray-300">Architecture design</th>
<th class="px-6 py-3 text-center font-semibold text-gray-900 border-b border-gray-300">Latency</th>
<th class="px-6 py-3 text-center font-semibold text-gray-900 border-b border-gray-300">Accuracy</th>
<th class="px-6 py-3 text-center font-semibold text-gray-900 border-b border-gray-300">Coordination</th>
</tr>
</thead>
<tbody class="divide-y divide-gray-200">
<tr>
<td class="px-6 py-4 text-gray-900">Lightweight orchestration</td>
<td class="px-6 py-4 text-center text-gray-700">~1.3×</td>
<td class="px-6 py-4 text-center text-gray-700">high</td>
<td class="px-6 py-4 text-center text-gray-700">high</td>
</tr>
<tr>
<td class="px-6 py-4 text-gray-900">Role-based pipelines</td>
<td class="px-6 py-4 text-center text-gray-700">5–30×</td>
<td class="px-6 py-4 text-center text-gray-700">medium</td>
<td class="px-6 py-4 text-center text-gray-700">medium</td>
</tr>
<tr>
<td class="px-6 py-4 text-gray-900">Simulation-style agents</td>
<td class="px-6 py-4 text-center text-gray-700">100×+</td>
<td class="px-6 py-4 text-center text-gray-700">low</td>
<td class="px-6 py-4 text-center text-gray-700">fails</td>
</tr>
</tbody>
</table>
</div>
<p class="text-sm text-gray-600">
Results from controlled experiments in MAFBench across multiple frameworks with identical LLM models and tasks.
</p>
</div>
</section>
<!-- Design Dimensions from Section 2 -->
<section class="py-20 px-4 sm:px-6 lg:px-8 bg-gray-50">
<div class="max-w-6xl mx-auto">
<h2 class="text-4xl font-bold text-gray-900 mb-12">Design Dimensions</h2>
<!-- Orchestration Design -->
<div class="mb-16">
<h3 class="text-3xl font-semibold text-gray-900 mb-4">Orchestration Design</h3>
<p class="text-lg text-gray-700 mb-6 leading-relaxed">
The execution model, scheduling overhead, and interaction semantics determine how agents are
coordinated. Results show that orchestration overhead can create latency multipliers ranging
from 1.3× to over 117× compared to direct LLM calls.
</p>
<div class="bg-white border border-gray-300 rounded-lg overflow-hidden mb-4">
<table class="w-full text-sm">
<thead class="bg-gray-100">
<tr>
<th class="px-6 py-3 text-left font-semibold text-gray-900 border-b border-gray-300">Orchestration type</th>
<th class="px-6 py-3 text-center font-semibold text-gray-900 border-b border-gray-300">Latency impact</th>
</tr>
</thead>
<tbody class="divide-y divide-gray-200">
<tr>
<td class="px-6 py-4 text-gray-900">Direct calls</td>
<td class="px-6 py-4 text-center text-gray-700">baseline</td>
</tr>
<tr>
<td class="px-6 py-4 text-gray-900">Graph workflows</td>
<td class="px-6 py-4 text-center text-gray-700">moderate</td>
</tr>
<tr>
<td class="px-6 py-4 text-gray-900">Simulation loops</td>
<td class="px-6 py-4 text-center text-gray-700">extreme</td>
</tr>
</tbody>
</table>
</div>
</div>
<!-- Memory Architecture -->
<div class="mb-16">
<h3 class="text-3xl font-semibold text-gray-900 mb-4">Memory Architecture</h3>
<p class="text-lg text-gray-700 mb-6 leading-relaxed">
The choice between accumulation and retrieval strategies fundamentally affects performance.
Accumulation leads to runtime blow-up and recall degradation as context grows. Retrieval-based
approaches maintain stable performance with bounded cost. Experiments show memory scores ranging
from 6.1% (accumulation) to 23.8% (retrieval).
</p>
<div class="bg-white border border-gray-300 rounded-lg overflow-hidden mb-4">
<table class="w-full text-sm">
<thead class="bg-gray-100">
<tr>
<th class="px-6 py-3 text-left font-semibold text-gray-900 border-b border-gray-300">Memory strategy</th>
<th class="px-6 py-3 text-center font-semibold text-gray-900 border-b border-gray-300">Accuracy</th>
<th class="px-6 py-3 text-center font-semibold text-gray-900 border-b border-gray-300">Cost</th>
</tr>
</thead>
<tbody class="divide-y divide-gray-200">
<tr>
<td class="px-6 py-4 text-gray-900">Retrieval</td>
<td class="px-6 py-4 text-center text-gray-700">stable</td>
<td class="px-6 py-4 text-center text-gray-700">bounded</td>
</tr>
<tr>
<td class="px-6 py-4 text-gray-900">Accumulation</td>
<td class="px-6 py-4 text-center text-gray-700">drops</td>
<td class="px-6 py-4 text-center text-gray-700">grows rapidly</td>
</tr>
</tbody>
</table>
</div>
</div>
<!-- Planning Interfaces -->
<div class="mb-16">
<h3 class="text-3xl font-semibold text-gray-900 mb-4">Planning Interfaces</h3>
<p class="text-lg text-gray-700 mb-6 leading-relaxed">
Free-form reasoning allows LLMs to plan naturally, while rigid schema constraints force
predefined formats that break LLM reasoning patterns. Results show accuracy changes ranging
from +15% (free-form iterative) to -30% (rigid schema), with runtime multipliers from 1.2×
to 30×.
</p>
<div class="bg-white border border-gray-300 rounded-lg overflow-hidden mb-4">
<table class="w-full text-sm">
<thead class="bg-gray-100">
<tr>
<th class="px-6 py-3 text-left font-semibold text-gray-900 border-b border-gray-300">Interface</th>
<th class="px-6 py-3 text-center font-semibold text-gray-900 border-b border-gray-300">Accuracy change</th>
</tr>
</thead>
<tbody class="divide-y divide-gray-200">
<tr>
<td class="px-6 py-4 text-gray-900">Free-form</td>
<td class="px-6 py-4 text-center text-gray-700">improves</td>
</tr>
<tr>
<td class="px-6 py-4 text-gray-900">Rigid schema</td>
<td class="px-6 py-4 text-center text-gray-700">−30%</td>
</tr>
</tbody>
</table>
</div>
</div>
<!-- Specialization Design -->
<div class="mb-16">
<h3 class="text-3xl font-semibold text-gray-900 mb-4">Specialization Design</h3>
<p class="text-lg text-gray-700 mb-6 leading-relaxed">
Procedural guidance and expert conditioning enable agents to specialize effectively.
Role clarity through procedural expert-guided conditioning improves F1 scores by up to 58 points
compared to generic agent designs.
</p>
<div class="bg-white border border-gray-300 rounded-lg overflow-hidden mb-4">
<table class="w-full text-sm">
<thead class="bg-gray-100">
<tr>
<th class="px-6 py-3 text-left font-semibold text-gray-900 border-b border-gray-300">Conditioning</th>
<th class="px-6 py-3 text-center font-semibold text-gray-900 border-b border-gray-300">F1 gain</th>
</tr>
</thead>
<tbody class="divide-y divide-gray-200">
<tr>
<td class="px-6 py-4 text-gray-900">None</td>
<td class="px-6 py-4 text-center text-gray-700">baseline</td>
</tr>
<tr>
<td class="px-6 py-4 text-gray-900">Expert-guided</td>
<td class="px-6 py-4 text-center text-gray-700">+58</td>
</tr>
</tbody>
</table>
</div>
</div>
<!-- Communication Topology -->
<div class="mb-16">
<h3 class="text-3xl font-semibold text-gray-900 mb-4">Communication Topology</h3>
<p class="text-lg text-gray-700 mb-6 leading-relaxed">
Connectivity patterns determine information flow and coordination limits. Fully connected
topologies enable success rates above 90%, while geometric topologies with isolated clusters
drop below 30%. The topology structure directly limits coordination capability.
</p>
<div class="bg-white border border-gray-300 rounded-lg overflow-hidden mb-4">
<table class="w-full text-sm">
<thead class="bg-gray-100">
<tr>
<th class="px-6 py-3 text-left font-semibold text-gray-900 border-b border-gray-300">Topology</th>
<th class="px-6 py-3 text-center font-semibold text-gray-900 border-b border-gray-300">Success rate</th>
</tr>
</thead>
<tbody class="divide-y divide-gray-200">
<tr>
<td class="px-6 py-4 text-gray-900">Fully connected</td>
<td class="px-6 py-4 text-center text-gray-700">>90%</td>
</tr>
<tr>
<td class="px-6 py-4 text-gray-900">Small-world</td>
<td class="px-6 py-4 text-center text-gray-700">medium</td>
</tr>
<tr>
<td class="px-6 py-4 text-gray-900">Geometric</td>
<td class="px-6 py-4 text-center text-gray-700"><30%</td>
</tr>
</tbody>
</table>
</div>
</div>
</div>
</section>
<!-- Why This Happens -->
<section class="py-20 px-4 sm:px-6 lg:px-8 bg-white">
<div class="max-w-6xl mx-auto">
<h2 class="text-4xl font-bold text-gray-900 mb-4">Why This Happens (And How to Fix It)</h2>
<p class="text-xl text-gray-700 mb-12 leading-relaxed">
Understanding the root causes is the first step to building better systems.
</p>
<div class="space-y-8">
<div>
<h3 class="text-2xl font-semibold text-gray-900 mb-4">Overhead Explodes</h3>
<p class="text-lg text-gray-700 leading-relaxed">
Framework orchestration introduces overhead at every interaction point. Sequential execution
models force agents to wait, while simulation-style loops create exponential call chains.
Architecture alone drives latency from 1.3× to over 117× with identical models.
</p>
</div>
<div>
<h3 class="text-2xl font-semibold text-gray-900 mb-4">Memory Bloats</h3>
<p class="text-lg text-gray-700 leading-relaxed">
Accumulation strategies grow context linearly with each interaction, leading to exponential
runtime cost increases. Retrieval-based approaches maintain bounded cost by accessing only
relevant information when needed. Memory structure, not context size, determines performance.
</p>
</div>
<div>
<h3 class="text-2xl font-semibold text-gray-900 mb-4">Rigid Planning Breaks LLM Reasoning</h3>
<p class="text-lg text-gray-700 leading-relaxed">
LLMs reason through natural language patterns. Forcing rigid schema constraints conflicts
with these patterns, causing formatting failures and reasoning breakdowns. Free-form
interfaces that match LLM capabilities improve accuracy, while rigid schemas reduce it by up to 30%.
</p>
</div>
<div>
<h3 class="text-2xl font-semibold text-gray-900 mb-4">Topology Limits Coordination</h3>
<p class="text-lg text-gray-700 leading-relaxed">
Communication topology determines information diffusion speed and coordination capability.
Isolated clusters prevent agents from reaching consensus, while fully connected networks
enable fast agreement. Topology structure alone can change coordination success from above
90% to below 30%.
</p>
</div>
</div>
</div>
</section>
<!-- Practical Implications -->
<section class="py-20 px-4 sm:px-6 lg:px-8 bg-gray-50">
<div class="max-w-6xl mx-auto">
<h2 class="text-4xl font-bold text-gray-900 mb-4">The Real-World Cost of Poor Design</h2>
<p class="text-xl text-gray-700 mb-12 leading-relaxed">
These architectural choices don't just affect benchmarks—they determine whether your system ships, scales, and succeeds.
</p>
<div class="bg-white border border-gray-300 rounded-lg p-8">
<ul class="space-y-6 text-lg text-gray-700">
<li class="flex items-start">
<span class="text-red-600 mr-4 text-2xl">•</span>
<div>
<strong class="text-gray-900">Cost blow-ups:</strong> 100× latency multipliers translate directly to 100× API costs.
A system that should cost $100/month can easily cost $10,000/month with poor architecture.
</div>
</li>
<li class="flex items-start">
<span class="text-red-600 mr-4 text-2xl">•</span>
<div>
<strong class="text-gray-900">Scalability ceilings:</strong> Poor orchestration prevents systems from handling increased load.
You hit walls that better architecture would never encounter.
</div>
</li>
<li class="flex items-start">
<span class="text-red-600 mr-4 text-2xl">•</span>
<div>
<strong class="text-gray-900">Unreliable behavior:</strong> Memory degradation and planning failures create unpredictable outputs.
Users lose trust when systems behave inconsistently.
</div>
</li>
<li class="flex items-start">
<span class="text-red-600 mr-4 text-2xl">•</span>
<div>
<strong class="text-gray-900">Debugging complexity:</strong> Architectural bottlenecks are difficult to identify and fix after deployment.
Prevention through good design is far cheaper than retrofitting.
</div>
</li>
</ul>
</div>
</div>
</section>
<!-- Transition CTA -->
<section class="py-20 px-4 sm:px-6 lg:px-8 bg-indigo-600">
<div class="max-w-4xl mx-auto text-center">
<h2 class="text-3xl font-bold text-white mb-4">Ready to Build Better Systems?</h2>
<p class="text-xl text-indigo-100 mb-8">
Learn how to make the architectural choices that actually matter.
</p>
<a href="architecture/index.html" class="inline-block px-8 py-4 bg-white text-indigo-600 rounded-lg font-semibold text-lg hover:bg-gray-100 transition shadow-lg">
→ Architecture Guide: Design Systems That Scale
</a>
</div>
</section>
<!-- Footer -->
<footer class="bg-gray-900 text-gray-300 py-12 px-4 sm:px-6 lg:px-8">
<div class="max-w-7xl mx-auto">
<div class="grid md:grid-cols-4 gap-8 mb-8">
<div>
<h3 class="text-white font-bold text-lg mb-4">MAFBench</h3>
<p class="text-sm">A unified benchmark for evaluating multi-agent LLM frameworks.</p>
</div>
<div>
<h4 class="text-white font-semibold mb-4">Resources</h4>
<ul class="space-y-2 text-sm">
<li><a href="paper.html" class="hover:text-white">Paper</a></li>
<li><a href="mafbench.html" class="hover:text-white">Benchmark</a></li>
<li><a href="https://github.com/CoDS-GCS/MAFBench" target="_blank" class="hover:text-white">GitHub</a></li>
</ul>
</div>
<div>
<h4 class="text-white font-semibold mb-4">Learn</h4>
<ul class="space-y-2 text-sm">
<li><a href="architecture/index.html" class="hover:text-white">Architecture Guide</a></li>
<li><a href="results.html" class="hover:text-white">Results</a></li>
</ul>
</div>
<div>
<h4 class="text-white font-semibold mb-4">Connect</h4>
<ul class="space-y-2 text-sm">
<li><a href="about.html" class="hover:text-white">About</a></li>
<li><a href="https://github.com/CoDS-GCS/MAFBench" target="_blank" class="hover:text-white">GitHub</a></li>
</ul>
</div>
</div>
<div class="border-t border-gray-800 pt-8 text-center text-sm">
<p>© 2026 MAFBench. Research by Concordia University.</p>
</div>
</div>
</footer>
<script src="scripts/main.js"></script>
</body>
</html>