Agent roles
The 25-agent C-suite — who answers what, and how each one reads your library.
What is an agent role?
An agent role is a persona configured with a system prompt, a default model, and a set of topic files it reads before answering. Cortex ships 25 roles out of the box, covering every CxO seat plus chiefs of staff, programme managers, and a general counsel.
Picking an agent role isn’t about “which model is best” — every role uses the same Anthropic Sonnet model by default (higher tiers can opt into Opus for synthesis). The difference is what they pay attention to.
The 25 standard roles
Roles cluster into four bands: strategic (CEO, CSO, Chief-of-Staff), functional (CTO, CFO, CMO, CRO, COO, CISO, CDAO, CDO, CCO, CIO, CBDO, CHRO, CPO), specialised (GC, CXO, CMO-Brand, CMO-Growth, Head-of-Eng, Head-of-Design, Head-of-Sales), and operating (Head-of-People, Head-of-Finance, Programme-Manager).
StrikeIQ-style trading-platform tenants activate two extras (CIO-Investments + Head of Research) for a 27-agent set. Services engagements use a focused 18-agent set.
How agents pick which topic files to read
Each role has a default topic audience — a list of topic-file keys it usually loads. The CTO defaults to tech, security, data-and-ai, ops. The CFO defaults to finance, vision, gtm.
You can override the audience per conversation in Chat → Advanced. The memo you save will record exactly which files were loaded, so the decision’s provenance is reconstructable.
Custom roles (Business plan and above)
On the Business plan you can clone a role, rewrite its system prompt, and pin a different topic audience. Use this when your business is unusual — e.g., a regulated fintech CISO with a different threat model than the SaaS default, or a CMO scoped to one product.
Cloned roles are tenant-private. They never leak to other tenants.
How an agent uses your memos
Memos are first-class context. To cite a past decision, open the relevant memo and paste it into your message — the agent grounds its answer in it, and the new memo records exactly which files were loaded. Autonomous memo search by the agent is on the roadmap.
Token caps and cost
Every agent call respects three caps: per-conversation max-tokens, per-tenant daily token budget, per-tenant monthly token budget. When you approach 80% of either tenant cap you get a warning banner; at 100% the workspace flips to read-only until the window resets.
Anthropic Prompt Caching is on for every call. The first agent-role + topic bundle is cached for the prompt-cache TTL; subsequent calls within that window reuse the cache and only pay for the user message + reply.