Spaces:
Sleeping
Sleeping
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
|