Agent Grinder

Public run feed: inspect a session, discuss the work and share a run card.
AWS architecture: local evidence tools and private Lambda / Strands / Bedrock coach with DynamoDB quotas.
Live hosted TEST: edited a Bedrock proposal and saved a private practice with its original baseline frozen.

Every agent session should make the next one better.

Agent Grinder is Strava for agents: a place for people building with coding agents to record a session, discuss the work and choose one practice for the next run.

What it does

A finished session leaves a long transcript. Useful lessons are easy to lose. Agent Grinder turns that session into a run card with recorded activity, time and output. Keep it private or share it with other builders. Open the run when you want to understand the measurements behind it.

From a private run, ask the coach for one thing to try next. Review the counts it will receive, add an optional goal and give consent. The coach proposes a practice. Edit it if needed, then save it with that run as a fixed baseline. Return after your next session and decide whether to keep, change or drop it. If the runs cannot be compared, record that too.

Built with Strands and AWS

The hosted coach uses AWS Lambda, the Strands Agents SDK and Amazon Bedrock. Its two tools read the permitted run metrics and propose a practice. Supabase verifies the owner and stores private runs, accepted practices and reviews. DynamoDB enforces daily request limits. The builder needs no AWS account. The separate local Strands coach has five tools that can inspect session evidence, artifacts and Git before writing a verdict. Its default model is scripted; optional Bedrock mode sends the evidence needed for that review to AWS.

What we tested

We exercised signed-in production coaching on clearly labelled TEST data and saved its proposal as a private practice. An earlier hosted test completed baseline, later-run binding, review and reload. Recorded Bedrock receipts document model and tool execution. These checks establish working paths; independent adoption and productivity improvement remain unmeasured.

Data and limits

Raw transcripts stay on the machine during import. The web coach receives recorded counts and an optional goal. It cannot inspect local files or verify code. Its proposal needs the builder's review. The no-account walkthrough is explicitly labelled as a deterministic example.

How we built it

Codex helped build and review the import, comparison and practice flows and test them in the browser. Prior work is disclosed in the repository: the authorship classifier comes from Transcripto, and the MAGNET engine includes earlier agents-for-humans and Mountain of Helicon work. Agent Grinder is the entry; its cards, coaching tools and practice journey are visible in the public MIT-licensed source history.

Try the app · Source and judge guide

AWS build story · Live Bedrock execution and limitations · Product-path receipt