The problem
Independent technical publications and solo creators spend an absurd amount of time looking for things worth covering. The work is repetitive but not mindless: discover a new tool, check whether it is genuinely new, find primary sources, inspect the repo/docs, reject duplicates and marketing fluff, judge whether the audience would care, then turn the survivors into something an editor can quickly approve or kill.
Most candidates are rubbish. The expensive part is not writing. It is the repeated research and triage before a human judgement is even worth making.
Toolglass Scout
Toolglass Scout is a Professional Agent built with the Strands Agents SDK for small technical publications, researchers, newsletter writers and creators who need to keep watch on a fast-moving field without spending their day refreshing feeds.
The agent works quietly in the background and handles the repetitive pipeline end to end:
- Scout selected public sources for newly released, obscure or unexpectedly useful software.
- Verify promising candidates against primary sources such as official documentation, repositories and release notes.
- Deduplicate them against material already covered or previously rejected.
- Judge fit against a small explicit editorial policy: novelty, usefulness, technical interest, evidence quality and likely reader interest.
- Build an evidence packet containing what it is, why it matters, important caveats, source links and the reason it passed the gate.
- Draft a short “Show & Tell” only for candidates that survive the research gate.
- Surface a human decision only when one is actually needed: approve, reject, investigate further, or hold.
The editor remains responsible for taste and publication. The agent owns the repetitive search, checking, deduplication and preparation that make that judgement possible.
Why an agent rather than another feed reader?
A feed reader gives you more things to read. Toolglass Scout is designed to give you fewer.
The useful behaviour is the sequence of autonomous work between discovery and interruption. It can follow links, gather evidence, compare a candidate with existing coverage, abandon weak leads without bothering the user, and retain enough history not to rediscover the same thing three days later.
That makes the human interruption meaningful: when Toolglass Scout surfaces something, there should already be a reason to care.
How Strands fits
The build uses the Strands Agents SDK to coordinate bounded specialist roles for discovery, verification, editorial fit and drafting. The orchestration is deliberately explicit rather than one giant prompt, so each stage can leave evidence and be tested independently.
Amazon Bedrock provides model reasoning. The target deployment is Amazon Bedrock AgentCore so the agent can run as a durable background worker with session/state boundaries rather than only as a local demo. AWS storage is used for compact candidate history, provenance and deduplication state.
The agent never auto-publishes merely because a model likes something. Publication remains a human editorial action.
The real-world test
Toolglass is an actual small computing magazine/project, so the hackathon build has a real workload rather than a synthetic demo. The success criterion is simple:
Can Toolglass Scout repeatedly find software the editor did not already know about, reject most of the noise, and produce a small number of evidence-backed candidates that are genuinely worth publishing?
A successful demo shows the entire path from discovery to verified candidate to human decision, including at least one candidate the agent rejects on its own.
Why it matters
The web has no shortage of information. The scarce resource is attention.
Small publications and independent creators rarely have research desks, but they are often the people most willing to cover strange, early or overlooked work. An agent that removes the repetitive scouting burden can let one person behave a little more like a small editorial team without replacing the part that actually requires taste.
Toolglass Scout is therefore not trying to automate an editor. It is trying to give the editor their time back.
Built With
- agentcore
- amazon-bedrock
- amazon-web-services
- github
- python
- strands-agents