How Nexard executes grounded agent workflows.

An inspectable architecture connecting isolated workspaces, keyword retrieval, server-side model routing, schema-validated structured decisions, human-in-the-loop approvals, and factual run evidence.

End-to-end execution flow

Nexard is engineered around deterministic boundaries, inspectable decision events, and human oversight. Below is the lifecycle of an agent execution in the private preview.

Available · Session auth

01 · User request & session auth

Requests authenticate via secure session cookies or platform headers. Anonymous executions are prohibited.

Available · Owner-scoped

02 · Isolated workspace boundary

Every database record (agent, source, version, run, approval) is strictly bound to the authenticated owner.

Available · Immutable snapshots

03 · Versioned agent instructions

Prompts and goals are versioned. Executions link to immutable configuration snapshots.

Available · Exact match

04 · Sandboxed keyword retrieval

Retrieval executes only within documents uploaded to the active agent. Excerpts are formatted as untrusted context.

Available · Anthropic SDK

05 · Server-side model routing

Server calls Claude 3.5/3.7 Sonnet using official Anthropic SDK with exponential backoff and timeout.

Available · Zod validation

06 · Structured decision schema

Model generates JSON with thought summary, answer, citations, and optional proposed actions validated against a strict schema.

Available · Sandboxed

07 · Sandboxed read-only tools

Tool permissions are enforced independently of model instructions. External writes, mutations, or sends are blocked.

Available · State machine

08 · Human approval checkpoint

Sensitive or consequential proposals halt execution in awaiting_approval state until explicitly authorized by the owner.

Available · Cloudflare D1

09 · Factual audit evidence ledger

Inputs, outputs, latency, token usage, and decisions are recorded in Cloudflare D1 APAC for observable inspection.

Why Claude for primary reasoning

Nexard integrates Claude as the primary reasoning layer because of its verified technical strengths:

Faithful source grounding: Claude exhibits industry-leading citation accuracy and resists hallucinating facts absent from uploaded knowledge excerpts.
Strict negative constraint adherence: When instructed never to promise refunds, collect banking details, or make unilateral commitments, Claude strictly complies.
Reliable structured output: Claude delivers valid JSON adhering strictly to Zod schemas without markdown artifacts or hallucinations.
Long-context comprehension: Effortlessly ingests extensive operational manuals and catalogs without context degradation.

Independent product disclaimer: Nexard is an independent product built on Cloudflare Workers and D1. All model calls execute via official server-side provider SDKs. Nexard is not affiliated with or endorsed by Anthropic.