Sign inGet started

Acts

Catch an agent going wrong, and set it right.

Guards hold a risky command before it runs. The steering hook nudges a looping agent back on course, or hands it to you. When the same mistake keeps coming back, Handz fixes the memory or instruction behind it, and you approve the change.

Get startedSee how it works
Works withEverything Tracez readsClaude Code hooksCodex hooksSandbox telemetry · coming
observerz.recurz.ailivesample data
Observerz mission control for an example 25-person team: 9 pieces of work running, 1 waiting on you; agents, pieces of work, model calls, spend, tokens and PRs opened in the last 24 hours; a timeline with one line per agent; the agent waiting on an answer; each piece of work with its helpers folded under it; spend per 15 minutes, by project, and a wall with one square per agent.
Helpers folded under the work that started themOne line per agent, coloured by projectSpend by project, not by folder

01 · Guards

Stop a risky command before it runs.

Guards check every command in Claude Code and Codex before it runs, such as a force push, a wide delete or a read of your credentials. Each guard starts in shadow and switches on only when its record shows it’s right.

observerz.recurz.ai/guardssample data
The Guards tab for an example team over 30 days: 92 receipts, 82 graded, 89% right, 2 rules ready to turn on; the Early stop guard in shadow, ready to turn on at 25 graded and 92% right, with right, wrong and precision per rule and model.
Ready to turn on once its record proves itEvery receipt graded right or wrong

02 · Nudges and take-over

Get a stuck agent moving again.

When an agent loops on the same failing test or quits early, the steering hook sends it a nudge written from what it was doing. If a nudge isn’t enough, the agent waits for you and you take over.

tracez.recurz.ai/sessions/S1sample data
A session’s loop finding: go test ./internal/refunds -run TestRefundWebhook failed 3 times in a row, then passed at step 45; the steps in between, and $0.24 lost to the loop.
The loop, its cost, and when it recovered

03 · Memory and instruction fixes

Fix the rule, not just the run.

When the same mistake keeps coming back across sessions, Handz finds the memory or instruction behind it and proposes an edit. You approve it, and it reaches every agent on every machine.

observerz.recurz.ai/fixessample data
The Fixes tab for an example team: 24 fixes to make in the last 7 days; counts of failing commands, loops and re-read files; the same Playwright failure in 8 pieces of work, what it printed, when it happened, and a change to try.
The same mistake, 8 pieces of workThe change that would stop it

How it works

From a stuck step to a fixed rule.

Handz watches the shape of the work, steps in, and fixes what caused it.

  1. Watch. Steps arrive from every machine, live.
  2. Catch. The same command fails again and again.
  3. Nudge. The steering hook sends a nudge, or hands the agent to you.
  4. Fix. A mistake that keeps coming back becomes an edit to the rule behind it.

Better together

Handz reads Tracez and feeds the rest

TracezWhere every session comes from. VerdictzTurns a finding into an eval, so the fix stays fixed. SteerzCaps spend at the call, before a loop can run up the bill.

04 · Acts

Who fixes it?

Six products answer it. Each one is useful alone, and each hands its work to the next verb in the loop.

  1. HandzGuards, nudges and take-over for agents going wrong.You are here
  2. FixezTurns errors and wrong results into fix PRs.Coming soon
  3. RefactorzPays down messy code with cleaner PRs.Coming soon
  4. AcceleratezFinds slow code in your traces and ships the faster PR.Open Acceleratez →
  5. OncallzAnswers the page on your machines and clouds, with evidence.Coming soon
  6. SquadzA swarm with a goal and a budget, grown where the work pays off.Coming soon

Early access

Let your agents run. Handz steps in when they go wrong.

Get startedEmail the foundersSee the API

Available now: guards in shadow, nudges from the steering hook, take-over, proposed memory and instruction fixes. Coming: sandbox telemetry, alert rules.