Context pointers: dynamic URL/doc references instead of static markdown

[louismorgner]

Problem

Agent context today is static markdown files baked into the agent definition. This works for simple instructions but breaks down when:

• Context needs to stay current (API docs, runbooks, internal wikis)
• Multiple agents share the same reference material (duplication + drift)
• Context is large and only conditionally relevant

Proposal

Add a context block to agent definitions that supports pointers — URLs, file paths, or doc references that are fetched/resolved at runtime rather than inlined statically.

context:
  - url: https://docs.stripe.com/api/charges
    cache: 1h
  - url: https://internal.notion.so/runbook-deploys
  - file: ./shared/style-guide.md

The runtime would fetch and inject these into the agent's prompt at session start (and optionally refresh mid-session for long-running agents).

Why this matters

Static context doesn't compound. If a doc updates, every agent referencing it is stale until someone manually copies the new version. Context pointers make agents automatically current — closer to how humans actually use reference material.

Open questions

• Caching strategy — TTL per pointer? Content-hash based?
• Auth for private URLs (internal wikis, Notion, Confluence)
• Size limits / truncation for large docs
• Should pointers support conditional inclusion (only fetch if agent hits a certain topic)?