Agentic feed summaries

Agentic AI News Agent

Coordinate source monitoring, clustering, summary generation, and verification without giving one model unrestricted access or publication authority.

AI news agentagentic workflowtool boundarieshuman reviewWebSub

Agent architecture

Bounded agency, visible states

The agent coordinates specialized tools instead of combining browsing, judgment, and publishing in one opaque loop.

State machine

Collecting, normalizing, clustering, ranking, summarizing, verifying, reviewing, and publishing are explicit states.

Least-privilege tools

Fetchers cannot publish; summarizers cannot browse; publishers accept only approved briefing objects.

Human gates

Topic and destination policies determine when review is mandatory and what can auto-publish.

Source control

A source registry controls reach

The agent should not improvise access to arbitrary websites.

Registry fields

Approved feed or API URL, host allowlist, parser profile, language, topic scope, cadence, credentials reference, access policy, owner, and health status.

Every run records the source configuration version.

Event handling

Scheduled checks and push events enqueue source-specific work with deduplication keys. The collector still validates and retrieves the source under policy.

Repeated events are idempotent and bounded by retry rules.

Editorial control

Review and publishing stay separate

A validated summary is not automatically an approved communication.

Generate draft

Use validated item and cluster records with explicit evidence IDs.

Verify independently

Check schema, facts, dates, quantities, entities, coverage, and source references.

Apply policy

Route high-risk topics and low-evidence results to mandatory review.

Approve or edit

Capture reviewer decisions as structured feedback and preserve the draft history.

Publish deterministically

Render the approved object, use an idempotency key, and record the delivery hash.

Feed Signals

Complete agent guide

Events, source registries, tools, clustering, verification, memory, security, review, and recovery.