V2 Runtime Control Plane

Stop Rebuilding Your Agent Control Plane.

Every quarter you rebuild the same infrastructure: task state, model routing, branch isolation, human review. HatiData V2 makes these platform entities — crash-safe, auditable, and queryable via SQL.

p99 entity write: <20ms

V2 ENTITY MODEL

11 First-Class Runtime Entities

Not log lines — authoritative, queryable entities with lifecycle, ownership, and lineage built in.

T
Task

Unit of work with metadata and lifecycle

A
TaskAttempt

Lease-based execution with heartbeat

M
ModelDecision

Which model was chosen and why

LlmInvocation

Tokens, cost, latency per call

ArtifactInstance

Output with confidence and validation

ReviewRequest

Human gate with evidence bundle

ReleaseDecision

Immutable approval record

W
WorkflowEvent

Append-only event stream

RecoveryAction

Automated retry or escalation

V
ArtifactValidation

Schema, policy, or human check

LineageEdge

Causal dependency in the DAG

EXPLAINBUNDLE

One API Call. Full Evidence Chain.

Every artifact carries its complete causal history — which task spawned it, which model decided, what evidence was evaluated, and who approved.

T

mortgage_review_batch_47

kind: compliance_review

A

att_8f2a…c901

agent: compliance-scanner-03

M

claude-sonnet-4-6

routing: cost_optimized

91%

1,847 tokens · $0.008 · 340ms

prompt_hash: sha256:a8f2…

risk_assessment_v2

confidence: 0.94 · validated

94%

BRANCH SAFETY

5 Visibility Modes for Speculative Execution

Every branch query declares how it sees the world. No surprises, no data leaks, no ambiguity.

Strict isolation — branch sees only its own writes.

const result = await hd.v2.branchQuery({
  branchId: "speculative-underwriting-23",
  visibility: "BRANCH_LOCAL",
  sql: "SELECT * FROM loan_decisions WHERE status = 'pending'"
});
// Only sees rows written in this branch — main is invisible

HUMAN GATES

Human Oversight as a Platform Entity

ReviewRequest is not a Slack message — it's a governed entity with evidence bundles, SLA enforcement, and immutable release decisions.

// Publish artifact with human gate
const artifact = await hd.v2.publishArtifact({
  attemptId: attempt.id,
  kind: "loan_approval",
  contentHash: "sha256:a8f2c901...",
  confidence: 0.72,
});

// Gate triggers automatically when confidence < 0.80
// Assigned to compliance team with 4-hour SLA
// → ReviewRequest created with full evidence bundle
1

Gate Triggers

Confidence below threshold or policy match → ReviewRequest created automatically

2

Evidence Accumulates

Full lineage chain, model decisions, validation results attached to the review

3

Decision Recorded

Approve or reject → immutable ReleaseDecision with rationale and timestamp

V2 RUNTIME SUBSTRATE

Three Pillars of Governed Autonomy

Control, transparency, and learning — as platform entities, not application features.

CONTROL

Governed Execution

Branch Contracts enforce visibility modes. Review Gates require evidence bundles before release. Every ReleaseDecision is immutable and auditable.

  • 5 Branch Visibility Modes
  • Human Gate Predicates
  • Immutable ReleaseDecision
  • Lease-Based Ownership
TRANSPARENCY

Causal Explainability

Mandatory lineage records every causal dependency as a DAG edge. ExplainBundle reconstructs the full evidence chain for any artifact in one API call.

  • Mandatory Lineage Edges
  • ExplainBundle API
  • Depth-Limited Graph Traversal
  • Orphan Artifact Detection
LEARNING

Optimization Substrate

Deterministic Replay re-runs past attempts with new models. Shadow Policies evaluate alternative routing without affecting production. Reward Records close the feedback loop.

  • Deterministic Replay
  • Shadow Policy Evaluation
  • Reward Records
  • Confidence Edge Tracking

OPERATOR VIEWS

Real-Time Runtime Visibility

SQL-queryable dashboards built on hd_runtime views. No custom monitoring stack required.

Active Leases
compliance-scanner-034:32
doc-extractor-070:48
underwriter-agent-010:12
fraud-detector-023:15
Pending Reviews
income_verification_01compliance@02:14:30
auto_approval_pkgunderwriting@00:45:12
high_value_loan_03risk@00:08:44
Confidence Distribution
≥0.9
847
0.7–0.9
312
0.5–0.7
156
<0.5
52

Ship Smarter Agents. Start in 60 Seconds.

Persistent memory. Isolated state. Verifiable reasoning. Through standard SQL.