Agents for Humans: building a Strava for agents
Turn coding sessions into run cards, try one practice and return with the next run. What we learned building the Strands and Bedrock coach.
Our first live Bedrock coach got the counts right and parts of the explanation wrong. Agent Grinder turns coding sessions into run cards and helps you choose one practice for the next session.
I wanted a Strava for agents. A place to post a run, see what someone else tried, and bring a useful idea into the next sitting with Cursor, Codex or another coding agent. The card is the entry point. You can keep it private, share it, discuss the work, or try a practice on your own baseline.
The coach gave us an early reason to be careful. It called a ratio a verification rate and read “outside this repository” as “in a different repository.” Its numeric verdict had passed our checks. The explanation still needed work.
We changed the export. The card now renders its summary directly from accepted measurements. The model's original explanation stays in the local report, labelled as interpretation. Proposed practices remain editable.
Five tools, one refused verdict
The transcript coach uses the Strands Agents SDK with read_run, check_claim, verify_artifact, git_evidence and write_verdict. It reads the session, looks for evidence in the same turn as each claim, inspects recorded artifacts and checks Git. The verdict tool can refuse counts that disagree with those results.
On 14 September, Claude Haiku 4.5 on Amazon Bedrock completed seven model requests and nine tool calls on our bundled test session. The SDK reported 17,123 input tokens and 1,550 output tokens. A follow-up through the product's coaching function recorded a refused verdict followed by an accepted one. Both receipts include the unedited output and its limits.
The default local mode runs the Strands loop with a scripted model and is labelled that way. Bedrock is opt-in. The opt-in Bedrock provider caps requests and output, with HTTP retries disabled.
The advice was to record what we already record
We then put a smaller Strands coach behind AWS Lambda, with two tools: read_run_metrics and propose_practice. Its first live replay suggested recording session metrics. Grinder already does that.
We added checks that reject metric-logging advice. A proposal must describe a workflow action, when to do it, and something the builder can inspect afterward. A later replay proposed pausing after edits to review each diff.
The signed-in flow now works on the hosted app. On a labelled test run, we reviewed the measurements, consented to sending them, received a live Bedrock proposal, edited it and saved a private practice. Grinder froze the original measurement revision as its baseline. The proposal needed editing; that step belongs in the product.
For the next session, the builder can return and choose whether to keep, change or drop the practice. We have tested that persistence flow. Independent adoption and improved productivity are still unproven.
What crosses the AWS boundary
The web coach receives recorded counts and an optional goal after consent. Lambda verifies the Supabase session, ownership and private visibility before reading an allowlist of fields. It does not read transcripts, titles, file paths or saved notes. The local transcript coach has a separate, explicit Bedrock opt-in.
DynamoDB reserves the web coach's daily allowances atomically: two requests per user and ten globally. A failed inference consumes its reservation. Lambda runs in us-east-1 with a US Bedrock inference profile. Supabase stores runs and practices; Vercel serves the app. AgentCore is not deployed.
The source discloses earlier MIT-licensed components from Transcripto, a session-analysis project, and agents-for-humans/MAGNET, our earlier coaching engine. Codex helped implement, review and test this build.
Try a private run and review what the coach proposes before accepting it.
App: https://agentgrinder.vercel.app
Code and setup: https://github.com/Morkeeth/agentgrinder
Demo: https://youtu.be/si5-DU7x-ho
Live execution receipt: https://github.com/Morkeeth/agentgrinder/blob/main/docs/BEDROCK-LIVE-2026-09-14.md
Follow-up receipt: https://github.com/Morkeeth/agentgrinder/blob/main/docs/BEDROCK-PRODUCT-PATH.md
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article