Black-box AI decisions
An answer alone cannot explain a decision. Reviewers need the tool activity, context, and human decisions that led to it.
Human-in-the-loop approvals, auditable thread history, and deterministic testing — built into the framework, not bolted on.
Tool calls, interrupts, and state transitions captured in the thread record
Each approval is written into the checkpoint beside the action it gated — the decision and the proposal are one record
Thread persistence preserves the full decision path for review by auditors and your compliance team
An answer alone cannot explain a decision. Reviewers need the tool activity, context, and human decisions that led to it.
Sensitive workflows need a clear place for a person to review a proposed action. That decision belongs in the flow, with the context to make it.
Compliance needs reproducible test evidence. Non-deterministic LLM calls make that nearly impossible without the right tooling.
Three libraries give your compliance team what they need — without slowing your engineering team down.
@threadplane/langgraphProduction agent state with first-class interrupt support. Every agent action can require human approval before execution. Durable thread persistence preserves the full record of every tool call and state transition.
Explore Agent@threadplane/renderApproval workflows rendered as structured UI — not chat messages. The agent proposes an action, renders a confirmation card, and waits for the human gate before proceeding.
Explore json-render@threadplane/chatDebug overlay shows every tool call, interrupt, and state transition. Your compliance team can review exactly what happened, when, and why — in a UI they can understand.
Explore ChatA focused Angular example, using the components and agent APIs behind this workflow.
audit-trail.component.ts — reading the thread record
export class AuditTrailComponent {
private readonly agent = injectAgent(REVIEW_AGENT);
// Runtime-neutral timeline: every checkpoint the thread passed through.
readonly checkpoints = computed(() => this.agent.history());
// Raw LangGraph ThreadState[], for the fields an auditor asks about.
private readonly raw = computed(() => this.agent.langGraphHistory());
// The decisions themselves, lifted out of the checkpoint values. Each row
// pairs what was proposed with what a human answered, and when.
readonly approvals = computed(() =>
this.raw()
.filter((state) => state.values?.['approval_result'])
.map((state) => ({
at: state.created_at,
action: state.values['proposed_action'],
decision: state.values['approval_result'],
})),
);
}audit-trail.component.html
<table class="audit">
@for (row of approvals(); track row.at) {
<tr>
<td>{{ row.at | date: 'medium' }}</td>
<td>{{ row.action.description }}</td>
<td>{{ row.decision.approved ? 'Approved' : 'Rejected' }}</td>
<td>{{ row.decision.reason }}</td>
</tr>
}
</table>
<p class="muted">{{ checkpoints().length }} checkpoints on this thread.</p>The agent proposes a destructive action, the graph pauses, and nothing runs until a human approves it.
Recorded from the live demo, which you can drive yourself.
Field report
Architecture, trade-offs and failure recovery
practical guidance
Angular and React implementation examples
both frameworks
Primary-source research and explicit API boundaries
free
Download the field report or start a pilot. Your compliance team will thank you.