Skip to content

Latest commit

 

History

History
648 lines (518 loc) · 38.2 KB

File metadata and controls

648 lines (518 loc) · 38.2 KB

Hive Agent Framework: Triangulated Verification for Reliable Goal-Driven Agents

The Core Problem: The Ground Truth Crisis in Agentic Systems

Modern agent frameworks face a fundamental epistemological challenge: there is no reliable oracle.

The dominant paradigm treats unit tests, execution results, or single-model evaluations as "ground truth" for agent optimization. Research reveals this assumption is critically flawed:

Assumed Ground Truth Failure Mode
Unit tests Binary signals lose quality nuance; coverage gaps allow overfitting; Goodhart's Law gaming
Model confidence (log-probs) Poorly calibrated; high confidence on wrong answers; optimizes for plausibility not correctness
Single LLM judge Hallucinated confidence; systematic biases; no calibration mechanism
Execution results Non-deterministic environments; flaky tests; doesn't capture intent

The consequence: agents optimized against these proxies become "optimizers of metrics" rather than "producers of value". They confidently generate code that passes tests but fails users.


Our Research Thesis: Triangulated Verification

Thesis: Reliable agent behavior emerges not from a single perfect oracle, but from the convergence of multiple imperfect signals.

We call this approach Triangulated Verification—borrowing from navigation, where position is determined by measuring angles to multiple known points. No single measurement is trusted absolutely; confidence comes from agreement across diverse signals.

The Triangulation Principle

                    ┌─────────────────┐
                    │  GOAL INTENT    │
                    │  (User's true   │
                    │   objective)    │
                    └────────┬────────┘
                             │
              ┌──────────────┼──────────────┐
              │              │              │
              ▼              ▼              ▼
       ┌──────────┐   ┌──────────┐   ┌──────────┐
       │Deterministic│   │ Semantic │   │  Human   │
       │   Rules   │   │Evaluation│   │ Judgment │
       └──────────┘   └──────────┘   └──────────┘
              │              │              │
              └──────────────┼──────────────┘
                             │
                             ▼
                    ┌─────────────────┐
                    │   CONFIDENCE    │
                    │  (Agreement =   │
                    │   reliability)  │
                    └─────────────────┘

Key insight: When multiple independent verification methods agree, confidence is justified. When they disagree, uncertainty should trigger escalation—not confident wrong answers.


The Three Verification Signals

Signal 1: Deterministic Rules (Fast, Precise, Narrow)

Programmatic checks that provide definitive verdicts for well-defined conditions:

  • Constraint violations (security patterns, forbidden operations)
  • Structural requirements (output format, required fields)
  • Known failure signatures (error types, timeout patterns)

Characteristics:

  • Zero ambiguity when they match
  • No false positives (if written correctly)
  • Cannot assess semantic quality or intent alignment

In Hive: EvaluationRule with priority-ordered conditions evaluated before any LLM call.

EvaluationRule(
    id="security_violation",
    condition="'eval(' in result.get('code', '')",
    action=JudgmentAction.ESCALATE,
    priority=200  # Checked first
)

Signal 2: Semantic Evaluation (Flexible, Contextual, Fallible)

LLM-based assessment that understands intent and context:

  • Goal alignment ("Does this achieve what the user wanted?")
  • Quality assessment ("Is this solution elegant/maintainable?")
  • Edge case reasoning ("What happens if input is empty?")

Characteristics:

  • Can assess nuance and implicit requirements
  • Subject to hallucination and miscalibration
  • Requires confidence gating

In Hive: HybridJudge LLM evaluation with explicit confidence thresholds.

if judgment.confidence < self.llm_confidence_threshold:
    return Judgment(
        action=JudgmentAction.ESCALATE,
        reasoning="Confidence too low for autonomous decision"
    )

Signal 3: Human Judgment (Authoritative, Expensive, Sparse)

Human oversight for high-stakes or uncertain decisions:

  • Ambiguous requirements needing clarification
  • Novel situations outside training distribution
  • Constraint violations requiring business judgment

Characteristics:

  • Highest authority but highest latency
  • Should be reserved for cases where automation fails
  • Provides ground truth for future automation

In Hive: HITL protocol with pause_nodes, requires_approval, and ESCALATE action.


The Triangulation Algorithm

┌─────────────────────────────────────────────────────────────────┐
│                     TRIANGULATED EVALUATION                      │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│  1. RULE EVALUATION (Priority-ordered)                          │
│     ┌─────────────────────────────────────────────┐             │
│     │ For each rule in priority order:            │             │
│     │   if rule.matches(result):                  │             │
│     │     return Definitive(rule.action)     ────────► DONE     │
│     └─────────────────────────────────────────────┘             │
│                          │                                       │
│                    No rule matched                               │
│                          ▼                                       │
│  2. LLM EVALUATION (With confidence gating)                     │
│     ┌─────────────────────────────────────────────┐             │
│     │ judgment = llm.evaluate(goal, result)       │             │
│     │                                             │             │
│     │ if judgment.confidence >= threshold:        │             │
│     │   return judgment                      ────────► DONE     │
│     │                                             │             │
│     │ if judgment.confidence < threshold:         │             │
│     │   return Escalate("Low confidence")    ────────► HUMAN    │
│     └─────────────────────────────────────────────┘             │
│                                                                  │
│  3. HUMAN ESCALATION                                            │
│     ┌─────────────────────────────────────────────┐             │
│     │ Pause execution                             │             │
│     │ Present context + signals to human          │             │
│     │ Human provides authoritative judgment       │             │
│     │ Record decision for future rule generation  │             │
│     └─────────────────────────────────────────────┘             │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Why This Order Matters

  1. Rules first: Cheap, fast, definitive. Catches obvious violations without LLM cost.
  2. LLM second: Handles nuance that rules cannot express. Confidence-gated.
  3. Human last: Expensive but authoritative. Only invoked when automation is uncertain.

This ordering optimizes for both reliability (multiple signals) and efficiency (cheapest signals first).


Goal-Driven Architecture: The Foundation

Triangulated verification answers "how do we evaluate?" But first we need "what are we evaluating against?"

Traditional agents optimize for test passage. Hive agents optimize for goal satisfaction.

Goals as First-Class Citizens

Goal(
    id="implement_auth",
    name="Implement User Authentication",
    description="Add secure user authentication to the API",

    # Multiple weighted criteria—not just "does it pass?"
    success_criteria=[
        SuccessCriterion(
            id="functional",
            description="Users can register, login, and logout",
            metric="llm_judge",
            weight=0.4
        ),
        SuccessCriterion(
            id="secure",
            description="Passwords are hashed, tokens are signed",
            metric="output_contains",
            target="bcrypt",
            weight=0.3
        ),
        SuccessCriterion(
            id="tested",
            description="Core flows have test coverage",
            metric="custom",
            weight=0.3
        )
    ],

    # Constraints: what must NOT happen (hard stops)
    constraints=[
        Constraint(
            id="no_plaintext_passwords",
            description="Never store or log plaintext passwords",
            constraint_type="hard",  # Violation = escalate
            check="'password' not in str(result.get('logs', ''))"
        ),
        Constraint(
            id="no_sql_injection",
            description="Use parameterized queries only",
            constraint_type="hard"
        )
    ]
)

Why Goals Beat Tests

Test-Driven Goal-Driven
Binary pass/fail Weighted multi-criteria satisfaction
Tests can be gamed Goals capture intent
Coverage gaps allow overfitting Constraints define hard boundaries
Silent on quality Success criteria include quality metrics

The Reflexion Loop: Learning from Failure

Research shows that iterative refinement beats expensive search. Reflexion (feedback → reflection → correction) outperforms MCTS in efficiency rank (accuracy/cost).

Worker-Judge Architecture

┌─────────────────────────────────────────────────────────────────┐
│                      REFLEXION LOOP                              │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│   ┌──────────┐         ┌──────────┐         ┌──────────┐        │
│   │  WORKER  │────────►│  JUDGE   │────────►│ DECISION │        │
│   │ Execute  │         │ Evaluate │         │          │        │
│   │   step   │         │  result  │         │          │        │
│   └──────────┘         └──────────┘         └────┬─────┘        │
│        ▲                                         │               │
│        │                                         ▼               │
│        │    ┌─────────────────────────────────────────┐         │
│        │    │  ACCEPT: Continue to next step          │         │
│        │    ├─────────────────────────────────────────┤         │
│        └────│  RETRY:  Try again with feedback        │◄─┐      │
│             ├─────────────────────────────────────────┤  │      │
│             │  REPLAN: Strategy failed, regenerate    │──┘      │
│             ├─────────────────────────────────────────┤         │
│             │  ESCALATE: Human judgment needed        │────►HITL│
│             └─────────────────────────────────────────┘         │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Feedback Context for Replanning

When a plan fails, the feedback loop provides rich context:

feedback_context = {
    "completed_steps": [...],      # What succeeded
    "failed_steps": [{             # What failed and why
        "step_id": "generate_api",
        "attempts": 3,
        "errors": ["Type error on line 42", "Missing import"]
    }],
    "accumulated_context": {...},  # What we learned
    "constraints_violated": [...]  # Hard stops triggered
}

This enables the planner to learn from failure rather than blindly retrying.


Uncertainty as a Feature, Not a Bug

Traditional agents hide uncertainty behind confident-sounding outputs. Hive agents surface uncertainty explicitly.

Four Levels of Capability

class CapabilityLevel(Enum):
    CANNOT_HANDLE = "cannot_handle"  # Wrong agent for this task
    UNCERTAIN = "uncertain"           # Might help, not confident
    CAN_HANDLE = "can_handle"         # Yes, this is my domain
    BEST_FIT = "best_fit"            # Exactly what I'm designed for

Graceful Degradation

High Confidence ──────────────────────────────► Low Confidence

┌──────────┐    ┌──────────┐    ┌──────────┐    ┌──────────┐
│ ACCEPT   │    │  RETRY   │    │ REPLAN   │    │ ESCALATE │
│          │    │  with    │    │  with    │    │  to      │
│ Continue │    │ feedback │    │ context  │    │  human   │
└──────────┘    └──────────┘    └──────────┘    └──────────┘
     │               │               │               │
     ▼               ▼               ▼               ▼
  Proceed      Learn from       Change          Ask for
              minor error      approach          help

Key principle: An agent that knows when it doesn't know is more valuable than one that confidently fails.


The Complete Picture

┌─────────────────────────────────────────────────────────────────────┐
│                         HIVE AGENT FRAMEWORK                         │
├─────────────────────────────────────────────────────────────────────┤
│                                                                      │
│  ┌─────────────────────────────────────────────────────────────┐    │
│  │                         GOAL LAYER                           │    │
│  │  • Success criteria (weighted, multi-metric)                 │    │
│  │  • Constraints (hard/soft boundaries)                        │    │
│  │  • Context (domain knowledge, preferences)                   │    │
│  └─────────────────────────────────────────────────────────────┘    │
│                              │                                       │
│                              ▼                                       │
│  ┌─────────────────────────────────────────────────────────────┐    │
│  │                      EXECUTION LAYER                         │    │
│  │  ┌──────────┐    ┌──────────┐    ┌──────────┐               │    │
│  │  │  Graph   │───►│  Worker  │───►│  Shared  │               │    │
│  │  │ Executor │    │   Node   │    │  Memory  │               │    │
│  │  └──────────┘    └──────────┘    └──────────┘               │    │
│  └─────────────────────────────────────────────────────────────┘    │
│                              │                                       │
│                              ▼                                       │
│  ┌─────────────────────────────────────────────────────────────┐    │
│  │                  TRIANGULATED VERIFICATION                   │    │
│  │                                                              │    │
│  │   Signal 1          Signal 2           Signal 3             │    │
│  │  ┌────────┐       ┌──────────┐       ┌─────────┐            │    │
│  │  │ Rules  │──────►│ LLM Judge│──────►│  Human  │            │    │
│  │  │ (fast) │       │(flexible)│       │ (final) │            │    │
│  │  └────────┘       └──────────┘       └─────────┘            │    │
│  │       │                │                  │                  │    │
│  │       └────────────────┴──────────────────┘                  │    │
│  │                        │                                     │    │
│  │                        ▼                                     │    │
│  │              Confidence from Agreement                       │    │
│  └─────────────────────────────────────────────────────────────┘    │
│                              │                                       │
│                              ▼                                       │
│  ┌─────────────────────────────────────────────────────────────┐    │
│  │                     REFLEXION LAYER                          │    │
│  │  • ACCEPT: Proceed with confidence                          │    │
│  │  • RETRY: Learn from failure, try again                     │    │
│  │  • REPLAN: Strategy failed, change approach                 │    │
│  │  • ESCALATE: Uncertainty too high, ask human                │    │
│  └─────────────────────────────────────────────────────────────┘    │
│                                                                      │
└─────────────────────────────────────────────────────────────────────┘

Roadmap: From Triangulation to Online Learning

Triangulated verification provides the foundation for a more ambitious capability: agents that learn and improve from every interaction. The architecture is designed to enable progressive enhancement toward true online learning.

The Learning Loop Vision

┌─────────────────────────────────────────────────────────────────────────┐
│                      ONLINE LEARNING LOOP                                │
├─────────────────────────────────────────────────────────────────────────┤
│                                                                          │
│                         ┌───────────────┐                                │
│                         │   EXECUTION   │                                │
│                         │  Agent acts   │                                │
│                         └───────┬───────┘                                │
│                                 │                                        │
│                                 ▼                                        │
│   ┌─────────────┐      ┌───────────────┐      ┌─────────────┐           │
│   │    RULE     │◄─────│ TRIANGULATED  │─────►│  CALIBRATE  │           │
│   │  GENERATION │      │  EVALUATION   │      │  CONFIDENCE │           │
│   │             │      └───────┬───────┘      │  THRESHOLDS │           │
│   └──────┬──────┘              │              └──────┬──────┘           │
│          │                     ▼                     │                   │
│          │            ┌───────────────┐              │                   │
│          │            │    HUMAN      │              │                   │
│          └───────────►│   DECISION    │◄─────────────┘                   │
│                       │  (when needed)│                                  │
│                       └───────┬───────┘                                  │
│                               │                                          │
│                               ▼                                          │
│                    Human decision becomes                                │
│                    training signal for:                                  │
│                    • New deterministic rules                             │
│                    • Adjusted confidence thresholds                      │
│                    • Signal weighting updates                            │
│                                                                          │
└─────────────────────────────────────────────────────────────────────────┘

Phase 1: Robust Evaluation (Current)

Status: Implemented

The foundation—triangulated verification provides reliable evaluation through multiple independent signals.

Component Implementation Purpose
Priority-ordered rules EvaluationRule with priority Fast, definitive checks
Confidence-gated LLM HybridJudge with threshold Semantic evaluation with uncertainty
Human escalation HITL protocol Authoritative fallback
Decision logging Runtime.log_decision() Record all judgments for analysis

What we can measure today:

  • Escalation rate (how often humans are needed)
  • Rule match rate (how often rules provide definitive answers)
  • LLM confidence distribution (calibration signal)

Phase 2: Confidence Calibration (Next)

Status: Designed, not yet implemented

Learn optimal confidence thresholds by comparing LLM judgments to human decisions.

@dataclass
class CalibrationMetrics:
    """Track LLM judgment accuracy against human ground truth."""

    # When LLM said ACCEPT with confidence X, how often did human agree?
    accept_accuracy_by_confidence: dict[float, float]

    # When LLM said RETRY, did the retry actually succeed?
    retry_success_rate: float

    # Optimal threshold that maximizes agreement while minimizing escalations
    recommended_threshold: float

    # Per-goal-type calibration (security goals may need different thresholds)
    threshold_by_goal_type: dict[str, float]

Calibration algorithm:

For each escalated decision where human provided judgment:
    1. Record: (llm_judgment, llm_confidence, human_judgment)
    2. If llm_judgment == human_judgment:
        → LLM was correct, threshold could be lowered
    3. If llm_judgment != human_judgment:
        → LLM was wrong, threshold should be raised
    4. Compute accuracy curve: P(correct | confidence >= t) for all t
    5. Set threshold where accuracy meets target (e.g., 95%)

Outcome: Agents automatically tune their confidence thresholds based on observed accuracy, reducing unnecessary escalations while maintaining reliability.

Phase 3: Rule Generation from Escalations (Future)

Status: Planned

Transform human decisions into new deterministic rules, progressively automating common patterns.

@dataclass
class RuleProposal:
    """A proposed rule learned from human escalation patterns."""

    # The pattern that triggered escalations
    trigger_pattern: str  # e.g., "result contains 'subprocess.call'"

    # What humans consistently decided
    human_action: JudgmentAction  # e.g., ESCALATE (for security review)

    # Confidence in this rule (based on consistency of human decisions)
    confidence: float

    # Number of escalations this would have handled
    coverage: int

    # Proposed rule (requires human approval before activation)
    proposed_rule: EvaluationRule

Rule generation pipeline:

┌─────────────────────────────────────────────────────────────────┐
│                    RULE GENERATION PIPELINE                      │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│  1. PATTERN MINING                                              │
│     ┌─────────────────────────────────────────────┐             │
│     │ Analyze escalated results for common traits: │             │
│     │ • Code patterns (regex over result.code)    │             │
│     │ • Error signatures (result.error types)     │             │
│     │ • Goal categories (security, performance)   │             │
│     └─────────────────────────────────────────────┘             │
│                          │                                       │
│                          ▼                                       │
│  2. CONSISTENCY CHECK                                           │
│     ┌─────────────────────────────────────────────┐             │
│     │ For each pattern, check human consistency:   │             │
│     │ • Did humans always decide the same way?    │             │
│     │ • Minimum N occurrences for confidence      │             │
│     │ • No contradictory decisions                │             │
│     └─────────────────────────────────────────────┘             │
│                          │                                       │
│                          ▼                                       │
│  3. RULE PROPOSAL                                               │
│     ┌─────────────────────────────────────────────┐             │
│     │ Generate candidate rule:                     │             │
│     │ • condition: pattern as Python expression   │             │
│     │ • action: consistent human decision         │             │
│     │ • priority: based on coverage + confidence  │             │
│     └─────────────────────────────────────────────┘             │
│                          │                                       │
│                          ▼                                       │
│  4. HUMAN APPROVAL (HITL)                                       │
│     ┌─────────────────────────────────────────────┐             │
│     │ Present rule to human for review:           │             │
│     │ • Show examples it would have caught        │             │
│     │ • Show edge cases for consideration         │             │
│     │ • Require explicit approval before active   │             │
│     └─────────────────────────────────────────────┘             │
│                          │                                       │
│                          ▼                                       │
│  5. DEPLOYMENT                                                  │
│     ┌─────────────────────────────────────────────┐             │
│     │ Add approved rule to evaluation pipeline:   │             │
│     │ • Shadow mode first (log but don't act)     │             │
│     │ • Gradual rollout with monitoring           │             │
│     │ • Automatic rollback if accuracy drops      │             │
│     └─────────────────────────────────────────────┘             │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Example learned rule:

# After 10 escalations where humans consistently rejected code with eval()
RuleProposal(
    trigger_pattern="'eval(' in result.get('code', '')",
    human_action=JudgmentAction.ESCALATE,
    confidence=1.0,  # 10/10 humans agreed
    coverage=10,
    proposed_rule=EvaluationRule(
        id="learned_no_eval",
        description="Auto-generated: eval() requires security review",
        condition="'eval(' in result.get('code', '')",
        action=JudgmentAction.ESCALATE,
        priority=150,  # Below manual security rules, above default
        metadata={"source": "learned", "examples": 10, "approved_by": "user@example.com"}
    )
)

Phase 4: Signal Weighting (Future)

Status: Conceptual

Learn which verification signals are most predictive for different goal types.

@dataclass
class SignalWeights:
    """Learned weights for combining verification signals."""

    # Per-goal-type weights
    weights_by_goal_type: dict[str, dict[str, float]]

    # Example:
    # {
    #     "security": {"rules": 0.7, "llm": 0.2, "human": 0.1},
    #     "ux": {"rules": 0.2, "llm": 0.6, "human": 0.2},
    #     "performance": {"rules": 0.5, "llm": 0.3, "human": 0.2},
    # }

Insight: For security goals, deterministic rules (pattern matching for vulnerabilities) are highly predictive. For UX goals, LLM judgment (understanding user intent) is more valuable. Learning these weights optimizes the evaluation pipeline for each goal type.

Implementation Priority

Phase Value Complexity Dependencies
Phase 1 High Done
Phase 2 High Medium Decision logging infrastructure
Phase 3 Very High High Phase 2 + pattern mining
Phase 4 Medium Medium Phase 2 + sufficient goal diversity

Recommended next step: Implement Phase 2 (Confidence Calibration) to enable data-driven threshold tuning. This provides immediate value (fewer unnecessary escalations) while building the dataset needed for Phase 3.


Research Contribution vs. Engineering Foundation

Layer Type Contribution
Triangulated Verification Research Novel approach to the Ground Truth problem; confidence from signal agreement
Online Learning Roadmap Research Architecture enabling agents to improve from human feedback over time
Goal-Driven Architecture Research + Engineering Goals as first-class citizens; weighted criteria; hard constraints
Confidence Calibration Research + Engineering Data-driven threshold tuning based on human agreement rates
Rule Generation Research Transforming human decisions into deterministic rules (closing the loop)
HybridJudge Engineering Implementation of triangulation with priority-ordered evaluation
Reflexion Loop Engineering Worker-Judge architecture with RETRY/REPLAN/ESCALATE
Graph Execution Engineering Node composition, shared memory, edge traversal
HITL Protocol Engineering Pause/resume, approval workflows, escalation handling

Summary

The Hive Agent Framework addresses the fundamental reliability crisis in agentic systems through Triangulated Verification and a roadmap toward Online Learning:

  1. The Problem: No single evaluation signal is trustworthy. Tests can be gamed, model confidence is miscalibrated, LLM judges hallucinate.

  2. The Solution: Confidence emerges from agreement across multiple independent signals—deterministic rules, semantic evaluation, and human judgment.

  3. The Foundation: Goal-driven architecture ensures we're optimizing for user intent, not metric gaming. The reflexion loop enables learning from failure without expensive search.

  4. The Learning Path: Human escalations aren't just fallbacks—they're training signals. Confidence calibration tunes thresholds automatically. Rule generation transforms repeated human decisions into deterministic automation.

  5. The Result: Agents that are reliable not because they're always right, but because they know when they don't know—and get smarter every time they ask for help.


References

  • Reflexion: Shinn et al., "Reflexion: Language Agents with Verbal Reinforcement Learning"
  • Goodhart's Law in ML: "When a measure becomes a target, it ceases to be a good measure"