Global Advisors built an internal agent platform that separates runtime orchestration from reusable skills, organisational memory, user memory and persona design. The goal was to move beyond isolated prompts and assistants toward durable, delegated AI coworkers that could participate in real work without losing context or crossing authority boundaries.

The work demonstrated that an agent is not a model wrapped in a chat interface. Useful agents need identity, tools, memory, retrieval, task state, permission scope, evidence, hand-off rules and a curation process.

Maturity: active internal use, with portability, evaluation, taxonomy and institutional operating disciplines still being strengthened.

Global Advisors needed an operating model in which:

• reusable capabilities were not trapped in one prompt;
• firm knowledge and personal context were distinct;
• agents could act for a named person within bounded scope;
• expert behaviour could be reused across interfaces;
• complex work could be delegated and recombined; and
• important learning survived model and runtime changes.

User memory

Durable context about a person is stored separately from general firm knowledge and temporary task state. Access is scoped so an agent aligned to one user does not automatically gain broad access to every user's memory.

Persona canon

Identity, role, tone and operating rules are represented as managed, portable assets rather than giant interface-specific prompts. Named agents, expert overlays and coach-style roles can share a common canon while retaining different authority.

• A live multi-agent runtime with a coordinating or curation role and delegated specialist work;
• shared repositories for skills, organisational memory, user context and personas;
• startup and workspace processes that materialise the right context for the active agent;
• scoped write access for user-aligned agents and broader curation authority for the central role;
• retrieval routes from thin memory briefs to deeper source collections;
• skills containing instructions, validation and deterministic support code;
• repeatable agent identities and role-specific behaviour across sessions;
• tooling for agents to work against real files, systems and workflows rather than only compose text;
• version control and review for changes to shared capability and canon; and
• a conceptual operating model connecting agents to workflow, knowledge, analytical and governance surfaces.

Delegation creates two problems that single-chat demonstrations can hide: context selection and responsibility. A specialist needs enough context to complete a bounded task, but not unrestricted access to everything known by the coordinating agent. Its output also needs to return with evidence and a clear status so the coordinator can integrate or challenge it.

We treated delegation as an operating contract:

• define the objective and boundaries;
• pass the smallest sufficient context;
• expose only appropriate tools and data;
• preserve the identity of the acting agent and represented user;
• return outputs with provenance or verification; and
• keep the coordinating agent responsible for synthesis and final judgement.

This resembles good organisational design more than prompt chaining.

Skills need more than prose

Complex work fails if a long instruction depends on perfect interpretation every time. Stronger skills combined judgement with schemas, scripts, examples and validation. Deterministic steps remained deterministic; models handled ambiguity and synthesis.

Tool access is delegated authority

An agent's ability to call a tool or write a file is an authority decision, not a convenience feature. The platform needed scope boundaries aligned to the user and task, with central curation separated from ordinary delegated work.