Research
Foundation models should be replaceable inference components. Whether a swap preserves execution is an empirical question. The hard problem is translating reasoning into correct actions across heterogeneous systems—ERP transactions, supply-chain state transitions, and governed treasury operations.
Workspace agents and retrieval systems support knowledge work. ERP execution additionally requires transaction controls, current state, recovery, and measurable completion. EGI isolates inference behind a versioned execution layer, then reruns the same job contracts when the model changes. Operational interchangeability is established by paired completion, safety, and recovery results—not by a shared API shape.
Public specifications for long-horizon jobs, paired model evaluation, auditable execution traces, and the runtime state machine behind controlled tool calling. The normative scoring contract is the ESE-1 specification.
ESE-1 · protocol v0.1
Long-horizon jobs, heterogeneous tools, scheduled state perturbations, and conjunctive pass/fail scoring.
Paired evaluation protocol
Hold the execution system fixed, change the inference model, and measure whether the completed job survives.
Trace schema · redaction rules
The minimum evidence required to reconstruct state binding, tool calls, writes, recovery, and completion.
Runtime specification
Transaction boundaries, policy gates, idempotency, verification, and resumable recovery.
Practical distinctions between retrieval, memory, orchestration, and execution—grounded in what each system can complete on a live ERP.
WMS · TMS · MES · logistics execution
Cash management · payments · banking APIs
NetSuite · SAP · Dynamics · posting
Agentic AI · autonomous agents · copilots
Agentic RAG · RAG vs agents
Context window · long-context LLM
EGI · AGI for the enterprise
Agent memory · context graph
Function calling · production agents
Workspace agents · scheduled tasks
Desktop agents · plugins
LangGraph · custom GPTs · build vs buy
Core principles
Why enterprise execution starts with standardized jobs, reusable workflow patterns, and production evidence from unlike tenant environments.
Read doctrine →09 · Sep 6, 2026
The practical test for an enterprise agent is whether it can complete a controlled transaction on a live ERP.
10 · Sep 6, 2026
Retrieval-augmented generation answers questions from a corpus. Agents have to complete jobs on live systems. Agentic RAG is still RAG.
11 · Sep 6, 2026
A million-token context window is still a transcript. Enterprise agents need current state, eviction, and a bind at the checkpoint—not a bigger paste.
12 · Sep 6, 2026
General intelligence in the enterprise is not AGI-as-chat and not a bigger foundation model. It is an agent that completes standardized functional work from Day 1.
16 · Sep 6, 2026
Neither is the DIY stack. They are inference surfaces and customization programs. Execution is a different object: jobs completed on systems of record.
Implementation notes for operators and engineers: checkpoint binding, symbolic execution, and action-level evaluation.
13 · Sep 6, 2026
A protocol note on the hippocampal gate—intent plus tool schema selects one slice on a lean index, binds it, executes, flushes.
14 · Sep 6, 2026
System 1 interprets ambiguous enterprise input. System 2 applies typed, rule-bound actions against production APIs.
15 · Sep 6, 2026
Seven execution dimensions, pass/fail on live systems, replication across unlike tenants. A model leaderboard cannot see any of this.
01 · Sep 6, 2026
They should be replaceable inference components. The claim is valid only when paired job evaluations survive the swap.
02 · Sep 6, 2026
Agents fail on stale, contradictory, and polluted state—not on a shortage of documents to retrieve.
03 · Sep 6, 2026
Orchestration is control flow. An execution graph is a learned path from intent to completion on real workflows.
04 · Sep 6, 2026
Demos succeed on small schemas and clean worlds. Production is the opposite—and function-specific.
05 · Sep 6, 2026
Bruce should not have to learn what purchasing, finance, warehouse, or operations are after you deploy.
06 · Sep 6, 2026
A model leaderboard cannot tell you whether the job completed. Production traces can.
07 · Sep 6, 2026
Permissions, approvals, observability, audit, and policy are not IT afterthoughts. They are how execution is allowed to exist.
08 · Sep 6, 2026
ERP is the canonical production surface. The product is correct action on the books—independent of the model underneath.
Enterprise AI agents complete consequential jobs on systems of record—not chatbots with plugins. ERP is the canonical workload; supply-chain execution adds physical state across WMS, TMS, and MES, while treasury adds governed cash movement across banking and payment systems. EGI’s product is an enterprise execution agent: intent through tool sequence, validation, recovery, and completion. Bruce is that agent.
No. Enterprise RAG and agentic RAG retrieve documents to answer questions. Agents have to change systems of record under current state, policy, and audit. EGI uses a context graph with just-in-time bind—not a vector corpus or session-wide preload.
No. A long context window is still a transcript. It does not update when enterprise state changes, and it cannot evict stale context. EGI binds a minimum slice at each checkpoint. Long context and RAG are both the wrong substrate for execution.
Enterprise general intelligence is operational capability that transfers across models, tenants, and functions while preserving function-specific controls. It arrives as pre-trained, benchmarked execution rather than a general chatbot or a customer-specific prompt project.
A copilot assists a person with a task; an execution agent is responsible for completing a defined job under explicit controls. ESE-1 scores outcome, tool correctness, state integrity, control adherence, recovery, and evidence conjunctively. The same frozen suite is rerun when the underlying model changes.
A knowledge graph stores entities and documents. EGI’s context graph is continuously refreshed enterprise state: addition, update, and eviction across structured and unstructured systems, bound at execution time so agents act on what is true now.
No. ChatGPT Work and Claude Cowork are agentic workspaces on a vendor’s foundation model. They finish documents, files, and research. An enterprise execution agent completes controlled transactions across ERP, supply-chain, and treasury systems under a context graph, execution graph, evals, and control plane. Bruce is that agent. The inference model is isolated as a replaceable component, and every change must pass the same job contracts.
DIY is the customization fallacy: a blank-slate agent, months of grounding, no standardized jobs, no cross-tenant benchmark. Frameworks give control flow. They do not give pre-trained execution. EGI ships the execution graph so you bind systems and policy rather than teaching the agent what a PO, journal, or close is.
Enterprise execution agent
The model is a dependency. Bruce is the agent that keeps the job correct—intent through completion on messy, heterogeneous ERPs. The serious work is posting, receiving, and closing.