Context Management & Reliability

1## CASE FACTS (Do not summarize — reference directly)
2
3| Field          | Value                                    |
4|----------------|------------------------------------------|
5| Customer       | John Smith                               |
6| Account ID     | ACC-12345                                |
7| Order          | #98765                                   |
8| Expected Price | $99.99 (promotion SUMMER2026)            |
9| Charged Price  | $150.00                                  |
10| Overcharge     | $50.01                                   |
11| Customer Since | 2019 (7-year tenure)                     |
12| Priority       | High (long-term customer + overcharge)   |
13
14## RULES
15- Always address customer as "Mr. Smith"
16- This case qualifies for immediate resolution
17- Refund amount ($50.01) is within $500 agent limit

1def should_escalate(context):
2    """Determine if we need human intervention."""
3    
4    # VALID escalation triggers
5    if context.customer_requested_human:
6        return True, "Customer explicitly requested human agent"
7    
8    if context.policy_gap_detected:
9        return True, "No policy covers this situation"
10    
11    if context.amount > AGENT_REFUND_LIMIT:
12        return True, f"Amount {context.amount} exceeds limit"
13    
14    if context.retry_count >= MAX_RETRIES:
15        return True, "Exhausted retry attempts"
16    
17    # INVALID triggers — DO NOT use these
18    # if context.sentiment == "negative":  # WRONG!
19    #     return True  # Sentiment != complexity
20    
21    # if context.model_confidence < 0.7:  # WRONG!
22    #     return True  # Self-reported confidence unreliable
23    
24    return False, None

1# Strategy 1: Scratchpad files for persistent state
2agent.run("""
3Before starting complex analysis:
41. Create a scratchpad file: progress.md
52. Record key findings as you discover them
63. Update progress.md after each major step
74. If context gets long, use /compact
85. After /compact, re-read progress.md to restore context
9""")
10
11# Strategy 2: Subagent delegation for verbose tasks
12coordinator = Agent(tools=[Task, read_scratchpad, summarize])
13coordinator.run("""
14For this codebase analysis:
151. Delegate file-by-file analysis to a subagent
16   (keeps verbose exploration out of coordinator context)
172. Subagent writes findings to scratchpad files
183. Coordinator reads summarized findings
194. Coordinator synthesizes final report
20""")
21
22# Strategy 3: Stratified metrics
23def track_accuracy(results):
24    """Track per-document-type, not just aggregate."""
25    by_type = {}
26    for r in results:
27        doc_type = r["document_type"]
28        if doc_type not in by_type:
29            by_type[doc_type] = {"correct": 0, "total": 0}
30        by_type[doc_type]["total"] += 1
31        if r["is_correct"]:
32            by_type[doc_type]["correct"] += 1
33    
34    # This reveals hidden failures per category
35    for doc_type, stats in by_type.items():
36        accuracy = stats["correct"] / stats["total"] * 100
37        print(f"{doc_type}: {accuracy:.1f}%")

1from dataclasses import dataclass
2from datetime import datetime
3from typing import Literal
4
5@dataclass
6class DataWithProvenance:
7    value: str | float | dict
8    source: str                    # "customer-db", "invoice-pdf", "web"
9    confidence: Literal["verified", "extracted", "inferred", "estimated"]
10    retrieved_at: datetime
11    agent_id: str                  # Which subagent provided this
12
13def resolve_conflict(data_points: list[DataWithProvenance]):
14    """When subagents disagree, use provenance to decide."""
15    
16    confidence_rank = {
17        "verified": 4,   # From authoritative source
18        "extracted": 3,  # Parsed from structured document
19        "inferred": 2,   # Derived from context
20        "estimated": 1,  # Best guess
21    }
22    
23    # Pick the most reliable source
24    best = max(data_points, key=lambda d: confidence_rank[d.confidence])
25    
26    # Log the conflict for audit trail
27    log_conflict(
28        chosen=best,
29        alternatives=data_points,
30        reason=f"Selected {best.source} (confidence: {best.confidence})"
31    )
32    
33    return best