Case Study · Agentic AI / Market & Content Research

Research Assistant

Weekly Agentic AI trend research, drafted into ready-to-use content, with a human always in control of what gets published.

RoleBuilder & Architect

ContextSolopreneur · Hackathon submission

DomainAgentic AI market & content research

TypePersonal automation tool

The problem

Anyone teaching or working in a fast-moving field faces the same weekly loop: check job boards, skim GitHub trending, read a stack of newsletters, try to tell real signal from noise — every week, from scratch, with no memory of what was already checked last time.

Who this is for

Independent AI trainers, educators, and consultants who need two different things from the same research — material to keep their own skills current, and concrete content to teach with — but don't have a research team to produce either.

Why it matters

Most research tools hand you information. This hands you a drafted LinkedIn post and outlined training modules — the gap between "here's what's happening" and "here's something you can publish or teach this week" is exactly what it closes.

How it works

Weekly trigger
Automation

GitHub Actions fires the workflow on a schedule, or on demand.
Orchestrator agent
Strands agent

Delegates to specialized sub-agents and synthesizes what comes back.
Job market & tooling trend sub-agents
Strands agents · parallel

Run in parallel — one researches hiring trends via search with a steering cap, the other tracks tooling trends via GitHub and RSS with the same steering guardrail.
Formatter agent
Strands agent · tool-less

Turns both sub-agents' findings into one structured digest — no tools attached, so there's no ambiguity about what it's allowed to do.
Digest archived & pushed
Automation

Saved to a digest archive, and pushed automatically and privately to the trainer via Telegram.
Trainer reviews
Human step

Edits or approves the drafted LinkedIn post before anything goes further.
Manual publish workflow
Automation · human-triggered

A separate GitHub Actions workflow, triggered only by the trainer, is what actually publishes.
Student Telegram channel
Delivery

Nothing reaches students without that manual trigger — the personal digest and the published content are deliberately separate paths.

What it actually does

Researches hiring trends and tooling trends in parallel, via two specialized Strands sub-agents delegated to by an orchestrator.
Remembers what it already reported — a dedupe memory layer means a weekly run surfaces only what's new, not the same items re-summarized.
Classifies every finding by signal strength — "strong" if corroborated across multiple sources, "early" if from just one — so it's clear what to teach confidently versus what to just watch.
Drafts content directly: a ready-to-edit LinkedIn post from the strongest finding, and 1–3 training-module ideas with learning objectives and key points.
Keeps a human in control of publishing: the digest reaches the trainer automatically and privately every week. Nothing reaches a student-facing channel without a separate, manually-triggered action.

Strands features used — and why each is there

Every one of these exists because a real bug or design need surfaced it during development, not as a checklist.

Feature	Why it's here
Multi-agent orchestration (Agent-as-Tool)	Delegation to specialized sub-agents, each with a narrow scope, instead of one agent doing everything.
Steering vended_plugins.steering	A real per-tool call budget with feedback the model can act on — replaced a hard-cutoff workaround after a sub-agent got stuck looping on a tool that couldn't do what it was trying to do.
Structured output Pydantic	The digest is a typed schema, not prose the model is asked to format.
Tool-less formatter agent	`structured_output()` internally disallows further tool calls — an agent that still has tools attached can try to use them anyway, which some model backends reject outright. A dedicated tool-less agent makes that failure structurally impossible.
ModelRetryStrategy	Real throttling protection, restored after being accidentally misconfigured while chasing an unrelated (tool-execution, not throttling) bug.
Native observability StrandsTelemetry	OpenTelemetry traces of every model and tool call, used throughout development to actually see the tool-call loop and delegation decisions.

Tech stack

Strands Agents SDK (Python) Groq AWS Bedrock (optional) GitHub Actions Telegram Bot API DuckDuckGo Search GitHub Search API RSS Feeds

Setup

cp .env.example .env   # fill in GROQ_API_KEY at minimum
pip install -r requirements_hackathon.txt
python hackathon_agent.py

For the scheduled + Telegram + approval-gated publish flow, see .env.example and the two workflows under .github/workflows/ in the repo.