Agent orchestration

Track finder.

Type
Personal multi-agent pipeline
Stack
Claude Managed Agents, Agent Skills, TypeScript, Vercel, Python
Status
Live, 2026

My Instagram feed finds better records than I have time to look up, so five Claude agents now do the digging for me.

I collect vinyl, and half my finds surface on Instagram these days. My routine was saving posts, then doing the Discogs and Spotify homework later. The backlog always won.

Now I share the post to a Telegram bot and it spawns a session on Anthropic's Managed Agents platform. An orchestrator routes it, a research agent identifies the track, and two agents write to my Discogs wantlist and a Spotify playlist in parallel. Cheap models do plumbing, expensive models do judgement, and every write runs through a deterministic Python skill.

Telegram capture bot
Webhook relay on Vercel
One agent session per shared post
Haiku orchestrator with a routing table
Instagram caption and image extraction
Parallel Discogs and Spotify fan out
Custom Python skills for every write
Vault held credentials, split tokens

The UI is a share sheet

A post shared to a Telegram bot, like sending it to a friend. The confirmation only arrives once a session actually exists.

Five agents, one direction

Agents talk in structured status blocks, never prose. Outlined chips are Haiku plumbing, filled chips are Sonnet judgement.

A relay that trusts nobody

Four gates before any work happens. The agent sandbox never holds the bot token; replies go back through a separate bearer token, valid for one chat.

A routing table, not a prompt

The orchestrator is a lookup running on Haiku. It forwards blocks verbatim and is forbidden to narrate, because every word gets re-read on every later turn.

Reading Instagram without an API

Sonnet curls the post as Googlebot, falls back to the image, and stops at eight tool calls. Known dead ends are listed in its prompt so they are never retried.

Fan out, then hold

Discogs and Spotify run in parallel. The first result waits for the second, so one batch reaches Telegram instead of a drip feed.

Scripts do the writes

The agent picks the pressing, a 1200 line Python skill does the wantlist write and explains its choice in one JSON object. Models never touch a write endpoint.

One message back

Routing fields stripped, results collapsed into a single message. Record wanted, track saved, homework done.

Orchestration turned out to be less about clever prompts and more about token economics. The prompts spend as many lines forbidding output as requesting it.

The design rule that emerged: models judge, scripts act. Packaging the scripts as Agent Skills keeps that boundary enforceable rather than aspirational.

Haiku routes and relays for pennies, Sonnet is saved for the two places judgement matters. Boring engineering, applied to agents.

Like what you see?
Tell me about your project
Grab a coffee