REUSABLE AGENT WORK

You steer the agent until it’s right.
What worked becomes a Play anyone can run.

Not a transcript, not a skill. The steps that worked, as code, at a link that runs on any machine. No expert needed.

curl -fsSL https://getrote.dev/playoffs/install.sh | sh

RUNS ONmacOSLinuxWindows via WSL

WORKS WITH

EXAMPLE SESSION
/play The agents spent $1,400 overnight. Why?
claude code › picking tools through play · billing api · logs · shell
AGENT✗compared model prices
YOU›it’s never the model. group the calls by tool.
AGENT✓fetch-invoice: 3,100 calls, every one a 404
YOU›cap retries at 3.
AGENT✓with the cap: $9
✓one tool, one bug, no retry cap$1,380 of $1,400
the job worked. keep what made it work as a Play?▸ YES, SAVE THE PLAY
✓Play mintedyour rules keptread-only
AFTER ONE GUIDED RUNPLAY READY
01 · A PLAY, AT A LINK
play.modiqo.ai/priya/overnight-bill
02 · ON A SCHEDULE
every morning at 07:00
03 · SHARED
#agent-ops · 6 people
THE PROBLEM

Running agents in production is hard.

THE SAME ASK · FOUR MORNINGS

/play The agents spent $1,400 overnight. Why?

  1. 01FLAKY
    ✓first run found the 404✗run again: it blamed the model

    The same request, run twice, takes two different paths.

  2. 02EXPENSIVE
    21ktokens to find the bug20kto find it again tomorrow

    Every run pays for discovery again, and the bill changes each time.

  3. 03UNGOVERNED
    ›read only, you said✗the agent edited retry.conf

    A rule for an agent is a post-it note. You cannot guarantee it holds.

  4. 04HARD TO SHARE
    ✓priya found it at 9 am✗on-call at 3 am starts from zero

    Good work stays in the chat where it happened. Everyone else starts from zero.

WHAT ROTE IS

Every call your agent makes is traced. The path that worked becomes a Play.

  • same path, every run
  • discovery paid once. the next run is a fraction of the tokens
  • the guard is in the file: read-only, credentials by name
  • a link anyone can run, on any machine

A Play is inspectable by design.

Write once. Run anywhere.The run that produced this file spent 21,400 tokens and took a wrong turn on the way. The Play does the same work in three declared steps for about 1,300, the same way every morning.

Move proven work beyond the frontier model.

  • discover once, at frontier prices
  • every repeat runs on a lower-cost model
  • one URI your team and the community can run, inspect, improve
≡×FRONTIER · ONCE
[.788] [.616] [.505] [.410]
[.400] [.805] [.265] [.273]
[.410] [.206] [.043] [.505]
[.273] [.505] [.765] [.665]
[.477] [.265] [.143] [.083]
[.087] [.043] [.095] [.900]

14,900 tokens · discovery
≡×PLAY
modiqo/hello
8 steps · read-only
digest 3f9c1a…

open modiqo/hello ↗
≡×REPEATS · ANY MODEL
run 07  haiku      ~300 ✓
run 08  gpt-mini   ~300 ✓
run 09  teammate   ~300 ✓
run 10  ci job     ~300 ✓

first run █████████ 14,900
every run █           ~300
MEASURED ON MODIQO/HELLO · 98% LESS REASONING · HOW WE MEASURED →
≡×FINGERPRINT · BEFORE THE RUN
adapter   openai-admin 2.3.0    ✓
endpoint  usage.completions     ✓
logs      ~/.claude · ~/.codex  ✓
scope     api.usage.read        ✓

GATE OPEN · run v1.0.0
≡×RUN
1  sessions   replayed   ✓
2  by tool    replayed   ✓
3  spend      replayed   ✓
 
4  retry cap  replayed   ✓

same steps · no reasoning
EXECUTION GATE · MATCH → REPLAY · CHANGED → REPAIR ONE STEP → V1.0.1

Recall stops when reality changes. It adapts, then remembers again.

  • before each run, the Play checks its fingerprint against the world
  • a match: it replays the known path, no reasoning
  • a change: only the changed step is reasoned again, the rest replays. the repair becomes v1.0.1, and the fingerprint moves with it
PUBLISHED PLAYS

Someone already solved it. Or you are about to.

≡×PUBLIC REGISTRY · EVOLVING
Watch expertise evolve.

Real Plays, versioned in public. Each one ran once at full cost; every run since follows the same path for a fraction. Read the contract, run it on your own credentials, publish the next version.

BROWSE THE PLAYS
≡×ROTE PLAYOFFS · 150 IDEAS
Find work worth teaching.

Research-backed starting points across APIs, browsers, and local commands. Pick one, hand your agent the brief, mint the Play.

PICK AN IDEA

Proven once. Adopted across the enterprise. Your organization's expertise, kept as Plays in its own private hub.

  • every steered run adds a Play to the org
  • every team runs it, on its own keys
  • private to the org. public only if you publish it
≡×DEVOPS
· · ·
waiting
≡×AGENTOPS
· · ·
waiting
≡×SECOPS
· · ·
waiting
≡×SALESOPS
· · ·
waiting
≡×/ACME INC
private hub · 6 plays · 4 authors
overnight-bill    ●●●● → public
release-brief     ●●●●
leaked-key-triage ●●●● → public
pipeline-hygiene  ●●●●
mcp-doctor        ●●●● ← adopted
cert-expiry-sweep ●●●●
PUBLISH →← ADOPT
≡×COMMUNITY
play.modiqo.ai · public
470 plays · anyone can run
token-tab         sidships
mcp-doctor        dotisacat
repo-fire-check   satianurag
overnight-bill    acme ★
leaked-key-triage acme ★
ONE HUB · EVERY TEAM ON ITS OWN CREDENTIALS · NOTHING REACHES THE COMMUNITY UNLESS YOU PUBLISH IT

Your first Play is one run away. Install Rote, ask your agent for something real, keep what worked.

≡×YOUR FIRST PLAY
  1. 01INSTALL
    curl -fsSL https://getrote.dev/playoffs/install.sh | sh
    Installs Play and Rote, detects your harness, starts guided setup.
  2. 02ASK
    $play build me a launch brief from GitHub, Linear, and Slack: what breaks if we ship Friday?
    Any harness. Claude Code, Codex, Cursor, Kimi, Hermes, or Pi.
  3. 03KEEP
    the job worked. keep what made it work as a Play?▸ YES, SAVE THE PLAY
    Play minted. Next runs: about 300 tokens instead of 14,900.