File size: 4,154 Bytes
df47251
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Agents System Design

## Overview

The agent runtime is a multi-agent, memory-aware RL orchestration layer for web extraction tasks. It supports:

- Single-agent and multi-agent execution modes
- Strategy selection (`search-first`, `direct-extraction`, `multi-hop-reasoning`)
- Human-in-the-loop intervention
- Explainable decision traces
- Self-improvement from past episodes

## Agent Roles

### 1. Planner Agent

Builds a plan before action:

- Goal decomposition
- Tool selection plan
- Risk and fallback path

### 2. Navigator Agent

Explores pages and search results:

- URL prioritization
- Link traversal policy
- Page relevance scoring

### 3. Extractor Agent

Extracts structured fields:

- Selector and schema inference
- Adaptive chunk extraction
- Long-page batch processing

### 4. Verifier Agent

Checks consistency and trust:

- Cross-source verification
- Conflict resolution
- Confidence calibration

### 5. Memory Agent

Manages memory write/read/search:

- Episode summaries
- Pattern persistence
- Retrieval ranking and pruning

## Execution Modes

### Single-Agent

One policy handles all actions.

Pros: low overhead, simple.
Cons: weaker specialization.

### Multi-Agent

Coordinator delegates work:

1. Planner emits execution graph
2. Navigator discovers candidate pages
3. Extractor parses and emits data
4. Verifier validates outputs
5. Memory Agent stores reusable patterns

Pros: modular, robust, scalable.
Cons: coordination overhead.

## Agent Communication

Shared channels:

- `agent_messages`: async inter-agent messages
- `task_state`: current objective and progress
- `global_knowledge`: reusable facts and patterns

Message schema:

```json
{
  "message_id": "msg_123",
  "from": "navigator",
  "to": "extractor",
  "type": "page_candidate",
  "payload": {
    "url": "https://site.com/p/123",
    "relevance": 0.91
  },
  "timestamp": "2026-03-27T00:00:00Z"
}
```

## Decision Policy

Policy input includes:

- Observation
- Working memory context
- Retrieved long-term memory hits
- Tool registry availability
- Budget and constraints

Policy output includes:

- Next action
- Confidence
- Rationale
- Fallback action (optional)

## Strategy Library

Built-in strategy templates:

- `search-first`: broad discovery then narrow extraction
- `direct-extraction`: immediate field extraction from target page
- `multi-hop-reasoning`: iterative search and verification
- `table-centric`: table-first parsing
- `form-centric`: forms and input structures prioritized

Strategy selection can be:

- Manual (user setting)
- Automatic (router based on task signature)

## Self-Improving Agent Loop

After each episode:

1. Compute reward breakdown
2. Extract failed and successful patterns
3. Update strategy performance table
4. Store high-confidence selectors in long-term memory
5. Penalize redundant navigation patterns

## Explainable AI Mode

Each action can emit:

- Why this action was chosen
- Why alternatives were rejected
- Which memory/tool evidence was used

Example trace:

```text
Action: EXTRACT_FIELD(price)
Why: Pattern "span.product-price" had 0.93 historical confidence on similar domains.
Alternatives rejected: ".price-box .value" (lower confidence 0.58), regex-only extraction (unstable on this layout).
```

## Human-in-the-Loop

Optional checkpoints:

- Approve/reject planned action
- Override selector/tool/model
- Force verification before submit

Intervention modes:

- `off`: fully autonomous
- `review`: pause on low-confidence steps
- `strict`: require approval on all submit/fetch/verify actions

## Scenario Simulator Hooks

Agents can be tested against:

- Noisy HTML
- Missing fields
- Broken pagination
- Adversarial layouts
- Dynamic content with delayed rendering

Simulation metrics:

- Completion
- Recovery score
- Generalization score
- Cost and latency

## APIs

- `POST /api/agents/run`
- `POST /api/agents/plan`
- `POST /api/agents/override`
- `GET /api/agents/state/{episode_id}`
- `GET /api/agents/trace/{episode_id}`

## Dashboard Widgets

- Live thought stream
- Agent role timeline
- Inter-agent message feed
- Strategy performance chart
- Confidence and override panel