Connector vs agent: the decision in one question
Does Copilot need to KNOW it, or DO something with it? Knowledge → connector (content joins the index; every Copilot surface benefits, zero runtime latency, no per-question API calls). Action or live state → API actions/agents (next concept). The most common extensibility mistake is building an agent to answer questions a connector would have fed to ALL of Copilot for free.
Engineering the pipe (what separates good connectors)
- Identity mapping quality: source principals -> Entra identities, per item. The data-plane concept's leak test is the acceptance bar; design time is where you earn it. Sources with no real ACLs (public wikis) get explicit everyone-mapping as a DECISION, documented.
- Schema thoughtfulness: which fields become searchable properties, labels for result types, activity signals if the source has them — ranking quality is built here, and it decides whether connector content ever actually WINS retrieval against native content.
- Incremental crawls: full crawl for the baseline, deltas on schedule — sources that can't express deltas force full re-crawls; budget the source-side load honestly.
- Result experience: url mapping so citations open the real item in the source (broken deep links kill trust — data-plane concept), icons and labels so users recognise provenance.
Build vs buy
Microsoft + partner gallery connectors cover the common systems; the SDK covers the rest (a connector is: auth to source, crawl loop, item transform, ACL mapper — a contained engineering artifact). Either way the review checklist and leak test are identical — provenance doesn't exempt gallery connectors from the quarterly re-test.
What to watch (proofs)
- The decision log: each extensibility request tagged connector/action/agent with the one-question rationale — the artifact that stops agent-for-everything culture.
- Crawl telemetry: items, deltas, errors, duration per sync — the pipe's health dashboard; error spikes precede stale-answer tickets.
- Retrieval share: are connector items actually appearing in citations? (Spot prompts that SHOULD hit them.) Ingested-but-never-retrieved means schema/ranking work, not more content.
- The standing leak test + link check per connector, quarterly — both screenshots into the connector's file (its 'MOT certificate').