The agent loop
Streaming, tool use, and parallel subagents across every provider — with tasks that survive compaction.
Moxie is the enterprise AI harness — the operator layer that runs digital workers across the business processes that power your company: Finance, Procurement, Supply Chain, Sales, Services, HR, Travel & Expense, and more.
The agent layer for your enterprise SaaS — the loop and the tools, the domain packs, the proof, and the surfaces you ship it on.
Streaming, tool use, and parallel subagents across every provider — with tasks that survive compaction.
Word, PowerPoint, Excel, PDF, and diagrams — plus a real browser and the desktop when no API exists.
OS-level sandboxing on every platform. Dangerous commands blocked, writes scoped to the workspace.
Markdown skills, MCP connectors, domain plugins — a dozen-plus in the box. Agents can author their own.
Scenario suites, hermetic benches, pass^k reliability — prove an agent works before it touches production.
Mac, iPhone, CLI, and web — thin clients over one gateway, white-labeled under your brand and sign-in.
Most agents forget every conversation. Moxie remembers in three governed tiers — what the agent carries, what the enterprise knows, and what it learns with a human in the loop.
Soul, user, project — each layer narrows what the layers below it can do.
Contract-governed data products, retrieved live from the system of record.
Institutional knowledge as an open-format graph; conflicts queue for human review.
Learnings wait as candidates until a person promotes them. Reversibly.
Enterprise AI doesn't stall on capability. It stalls because nothing decides what an agent has earned the right to do. Moxie ships that operating system.
Scenario evals with in-situ judges, hermetic benches on ephemeral infrastructure, pass^k reliability — a task passes when it passes every time.
Real work, no writes. Gated actions are captured as fully-specified proposals and graded against the system of record. Divergence is signal, not failure.
The action, not the agent, is the unit of trust. Reads may run free while writes to the system of record still need a person. Promotion is human-ratified; demotion is automatic.
Every deny, decision, and interrupt lands in a tamper-evident ledger. Recurring corrections become regression evals and staged learnings — the substrate for specialized intelligence.
Every run emits a receipt with evidence, and a blind verifier checks the claims before success is counted. Failed verification demotes the result — honestly.
"The path from promising pilots to mission-critical work is not unrestricted autonomy. It is bounded authority, earned through evidence."
The Project is the unit of collaboration — people, named agents, and agent teams sharing one home, one queue of work, one scoreboard. The calls that matter stay human; everything else keeps moving.
Persistent teammates, not throwaway sessions — each with its own memory, audit log, KPIs, and a driving record that decides what it may do alone.
A coordinator and roster share a scratchpad and a budget. Outcomes roll up across members, so the team is accountable as a unit.
Some calls are never the model's — plan elections, attestations, steward merges. They render as inline cards and block until a person answers. In Auto mode too.
A clean split: the Moxie harness runs the loop; your runtime owns identity, tenancy, audit, and policy. That separation is what makes Moxie adoptable — it runs on the control plane you already operate, rather than asking you to stand up another one. Every layer is provider-agnostic; nothing holds your data captive.
Moxie runs on the control plane you already operate. No new identity system to stand up, no data to move, no parallel stack to run beside the one you have.
OIDC · SCIM 2.0 · RBAC · scoped access tokens
Identity, secrets, policy, observability, and audit stay exactly where they are today. Moxie consumes them through typed seams instead of replacing them.
Vault · Splunk HEC · Syslog · Datadog · OTLP
Five trait seams keep the harness portable across runtimes. Swap the sandbox, the storage, or the secrets backend without touching the agent loop.
SandboxClient · SessionStore · IdentityVault · MemoryStore · ObservabilitySink
An agent earns broader autonomy only when its context, evidence, controls, and outcomes have earned it. Moxie is what that looks like when you actually build it.
I'm Eric Du. I've spent my career building enterprise platforms — most recently as SVP, Chief Architect and Head of Data and Platform at OpenText, and before that as CTO of SAP SuccessFactors.
Moxie is what I am building during my first career break. Instead of writing another point of view on where enterprise AI is going, I am putting the ideas, experience, and architecture into a working product — and learning what the argument really costs when it has to run.
The hard part is not making an agent capable. It is making one accountable: memory that spans a company rather than a chat; teams of agents that map to real business functions and can show what they accomplished; and a trust model where autonomy is earned task by task, grounded in evidence, and can be reduced when that trust is no longer justified.
Moxie is built by one person, heavily assisted by AI coding agents. That is not a footnote — it is part of the argument. The leverage available to a single builder today is the same leverage Moxie is designed to bring to an enterprise.