The Incumbents Are Coming — Loosely Coupled, Not Locked In

That AI makes systems of record more valuable, not less, because incumbents can control data/action access and push their own agents forward — but the customer's job to be done is bigger than any one record, leaving room for vertical AI-native startups that win through focus and a learning loop.

A partnership between Salesforce and Anthropic letting a user work with Salesforce CRM data and actions from inside Claude without opening Salesforce, with Claude as the front door while Salesforce still owns the underlying CRM data and actions.

Retrieval assistants, Process agents, Policy agents, and Principal agents — distinguished by how much autonomy and judgment each requires.

That each application may describe the same customer, contract, or transaction differently — pulling information across applications, even via MCP, has latency and no shared representation of the same real-world thing.

A shared Turtle-file knowledge graph in this repository that a dozen-plus independently-vendored AI agent harnesses all read and write, using one common vocabulary and globally dereferenceable IRIs — a concrete, running example of loose coupling.

By reading the same files — core.ttl, preferences.ttl, entities/*.ttl, howto/*.ttl, sessions/*.ttl — every other harness already reads: zero bespoke per-agent integration code, zero schema negotiation.

SPARQL federated queries (SERVICE) join across systems by shared IRIs instead of requiring N bespoke API connectors, one per data source.

Provenance (schema:author, schema:dateModified, PROV-O-style actedOnBehalfOf) comes for free from the graph model, instead of being re-implemented as custom audit logging inside every application.

A playbook or threshold rule set expressed once — as SHACL shapes, or schema:HowToStep sequences — is reusable across any record shaped by the same vocabulary, instead of being a closed model, like Docusign's Iris, that only ever applies inside one vendor's own schema.

A job that crosses applications and companies can be described once, as a graph of dereferenceable IRIs, and referenced FROM every system that touches it, rather than re-keyed by each system's private primary key.

Loose coupling was never meant to solve everything by itself — it's an enabling property, not a fixed algorithm. It's what lets you pick the right horse for a given part of the job: Harvey's manufactured curriculum (via its model Tenet) is one horse, changing model weights for judgment on ambiguous cases. The same graph substrate agent-rdf-memory runs on already saddles a second horse — a Beyond LLM Wiki pattern where tacit and explicit knowledge feed structured HowTo/Action entities that every future agent turn recursively re-consults, no training run required. Horses for courses, not a competition.

It post-trained an open-weight model using synthetic data, public legal data, and human-expert-created data, creating roughly 1,750 legal-task environments simulating partner-assigned matters, each with an expert rubric of about 50 specific criteria.

Salesforce's CRM (via Claudeforce), Docusign's Iris, Atlassian's Rovo, and Klaviyo's Composer.

Yes, at the in-context/behavioral layer: mistakes corrected in any session, across any of the dozen-plus model vendors sharing the graph, get written once as a schema:HowToStep and re-consulted by every future agent task-processing turn before it acts — a Beyond LLM Wiki pattern blending tacit knowledge (session corrections) and explicit knowledge (skill documentation). It does not change model weights the way Harvey's Tenet does; it changes what any agent reads before acting.