ContextWorks blog
What an agent needs to answer well.
Agents, language models, security, governance and semantics. How AI tools get trusted Salesforce context, and what goes wrong when they do not.
Introducing ContextWorks: Your CRM Data, Agent-Ready
Curate which fields an agent sees, explain what they mean, enforce permissions per user, and deliver it to Agentforce, Claude and ChatGPT from one definition.
A Short Glossary
The terms we use for making CRM data agent-ready, defined in one place.
Versioned, With Rollback
Change what an agent knows while people are using it. Versions, activation comments, one-action rollback, and change logs, on the Flow model admins already know.
One Definition, Every Surface
Agentforce, Claude, ChatGPT, Prompt Builder, Flow, Apex — one set of definitions behind all of them. Add or switch AI tools without rebuilding context.
Same Tool, Same Record, Two Users, Two Payloads
How ContextWorks shapes agent context with domains and enforces security with Salesforce permissions. Two users, one record, two correct payloads.
Four Pieces: Entities, Domains, Briefs, Skills
Entities carry the records. Briefs carry the institutional knowledge. Skills carry the procedure. Domains compose them for an agent.
What an Entity Is
Fields, order, rollups, filtered related lists, computed signals and instructions — assembled live, per user. Here's what an agent receives.
Ask Five People What "Open Deals" Means
An agent asked for open deals guesses IsClosed = false. See why business vocabulary needs to be written down once, not re-guessed per query.
Filtering the Answer Doesn't Make It Safe
Trimming an agent's results after the query leaves the model holding data the user can't see. Learn why enforcement belongs in the query itself.
Your CRM Wasn't Built for Agents to Read
Hundreds of fields, no semantics. Here's why connecting AI to Salesforce either gets refused or gives wrong answers, and what the fourth option looks like.
Your Agent Can Reach Every Field. That's the Problem.
Giving an agent every field makes it confidently wrong and quietly unsafe. See why curated, permission-correct context beats wider access.
Nothing published on this topic yet.
