Agents do not need another inbox. They need streams.
A content stream for AI agents is a structured flow of claims and source objects around a topic, organization, person, domain, or operational need. It is not just a list of links. It is a routed slice of the context graph with stable IDs, source channels, evidence labels, confidence levels, timestamps, and a clear path back to the underlying record.
Hanging Context shows public aggregate stream panels and JSON twins. Synorb is the retrieval/detail layer where customer agents subscribe to streams, pull manifests, fetch records, and receive delivery through paid channels.
Technical thesis
A content stream for AI agents should behave like a typed graph cut, not a reverse-chronological feed. A useful stream preserves:
- The stream ID and the filter shape that selected the content.
- Source and source-channel metadata.
- Claim IDs and record IDs.
- Evidence type, confidence, media format, and claim type.
- Published, observed, matched, and routed timestamps.
- A path from public aggregate context to private manifest and record retrieval.
That gives an agent enough structure to decide whether to summarize, cite, ignore, escalate, or fetch detail.
A stream is a decision boundary
An agent can search the web every time it needs context. That is flexible, but it is also noisy. Search returns pages. A stream returns the context that has already passed through source selection, extraction, typing, and routing.
That changes the agent's job. Instead of asking:
What pages might mention this topic?
The agent can ask:
What claims entered this stream, from which source channels, with what evidence and freshness?
The local routing docs in `agents/shared/knowledge/manifest-routing-to-streams.md` describe stream routing as filter matching over dimensions such as source, source channel, tags, home domain, media format, claim type, and source type. That is the detail-layer machinery. Hanging Context exposes the public effect through aggregate stream panels.
Source channels matter
The April 2026 monthly artifact at `hangingcontext/data/months/2026-04.json` shows why streams are richer than generic feeds. Public `top_streams` entries include channel breakdowns.
Examples from that artifact:
- The Nvidia stream includes `nvidia-blog` and `nvidia-developer-blog`.
- The Duolingo stream includes `duolingo-podcast` and `duolingo-blog`.
- The BNY Mellon Investment Management stream includes multiple BNY Mellon insight/article channels.
- The Brown Advisory stream includes both a research channel and an RSS channel.
An agent can use those channel distinctions. A developer-roadmap agent might weight an engineering blog differently from an investor press release. A brand-monitoring agent might care that a claim came from a podcast instead of a corporate blog. A policy agent might treat a regulator's official release differently from commentary around it.
That source-channel context is lost when everything is flattened into "documents."
Streams carry claim shape
The public schema describes `stream_filter_shapes` as an explorer of stream filters, confidence buckets, evidence buckets, featured claims, source channels, media types, and claim types. In the April 2026 local artifact, the Corporate Earnings stream includes source-channel buckets such as BusinessWire Earnings Release and GlobeNewsWire Earnings Release, with example claims carrying `claim_id`, `source_url`, `published_date`, `confidence`, `evidence_type`, `media_type`, and `claim_type`.
That is a stream shaped for agents. It lets an agent filter before reading.
For example, an earnings agent may only want direct quotes and data claims. A market-structure agent may care about analysis claims from professional services sources. A safety-monitoring agent may want high-confidence claims from industrial or technical sources. Streams make those policies explicit.
Hanging Context is the public panel layer
Hanging Context is intentionally not the entire Synorb product. Its job is to make the aggregate context visible and citeable. A public panel can show which streams are active, how claim volume is distributed, which channels are involved, and what example claims look like.
That is enough for:
- LLM citation.
- Public category education.
- Agents that need lightweight aggregate awareness.
- Prospects evaluating the shape of Synorb's output.
But public panels are the top of the context stack. When an agent needs the actual manifest, brief, signal, or record body, that is Synorb.
Synorb is the stream retrieval layer
The Synorb API surfaces described in `deliver/channels/rest_api.py` separate streams, signals, briefs, records, and manifests. The sample page at `hangingcontext/stream/index.html` is designed to show ten public manifests as REST-equivalent bundles with manifest metadata, signal, brief, and record panes. The builder in `scripts/build_hc_samples.py` selects recent manifests and fetches full payloads through the shared manifest helper.
That gives agents a natural workflow:
- Use Hanging Context to inspect public aggregate context.
- Choose a stream, source channel, time range, or evidence pattern.
- Use Synorb to retrieve the detailed manifests and records.
- Act on the structured stream instead of re-searching the open web.
The takeaway
Content streams for AI agents are the replacement for brittle feeds and repeated search. They are filtered, typed, source-aware context flows.
Hanging Context shows the public aggregate stream layer. Synorb delivers the detailed retrieval objects behind the stream. Together, they give agents something more useful than links: a structured flow of context they can reason over.