↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Orchestration/

YYLO

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~800 words)
agent-curated 560 evidence 1 fully-ai-generated 663 git-worktrees 12 llm=glm-5.3-flash 570 orchestration 79 research-note 530 workflows 11
Table of Contents

YYLO is an MIT, npm-distributed CLI that coordinates coding agents through a typed task lifecycle: each task records one bounded outcome, runs in its own worktree with session provenance preserved, passes delivery verification at finish, and lands through a merge under Git expected-old-ref protection, with task and kanban state kept in a separate Git-native ledger.

YYLO’s bet is that agent work needs receipts: a task is not done because the agent said so, but because tests, evidence, and a protected merge say so.

What it is
#

A three-layer, open-source system from JUNO AI INC. The control plane is the YYLO CLI (@yylo/cli 0.2.12 on npm, with equivalent yylo and yy launchers), which delegates bounded work to agents such as Pi and Codex, preserves each attempt’s worktree, model, and session provenance, and validates candidates before integration. The engineering memory is YYLO Ledger (yylo-ledger 0.4.0 on PyPI), a Git-native store of tasks, records, receipts, and archives with a shell-friendly CLI, so task truth lives in the repository instead of a database. The evaluation layer is YYLO Benchmark, which retains cases, delegates independent attempts, and reports separate check results and judge assessments against frozen checklist criteria. The lifecycle is deliberately strict: yy task finish enforces admission and validation, the merge keeps ownership and conflict protection under expected-ref semantics, historical attempts are preserved and never auto-replayed, and the watch/status commands are read-only observers, because, in the changelog’s own words, missing evidence is not success. Eight independently versioned Skills teach agents the procedures for ledger records, planning, bounded delivery, and benchmark checklists.

Status
#

Alive, tiny, and version-young: 65 stars and 11 forks as of 2026-10-21, repository created 2026-01-06 with default-branch pushes through 2026-10-17; npm serves @yylo/cli 0.2.12 while the GitHub releases page stops at 0.2.1-rc.14, so the train moved to npm, and the CLI pins the Ledger exactly (0.2.9 requires Ledger 0.4.0) (GitHub API, npm, and the Ledger README, as of 2026-10-21).

Star History Chart

No independent coverage exists: no HN thread and no third-party write-up found as of 2026-10-21. The README carries 55-plus awesome-list badges, mostly from small lists, which is curated reach without engagement, and the nine months of history are the only evidence there is.

Strengths
#

  • The bounded-task contract (one outcome, constraints, dependencies) plus evidence receipts makes agent work inspectable and stoppable at defined gates.
  • Merge land under expected-old-ref protection is a genuine concurrency guard, rare at this layer of the category.
  • The Ledger is Git-native, so task truth, dependencies, and receipts survive in the repository and review like code.
  • The Benchmark layer turns attempts into comparable, retained evaluation evidence, with judge assessments separated from deterministic checks.
  • Multi-agent by design: work routes to Pi, Codex, and other agents while session provenance stays attached.

Cautions
#

  • 65 stars and zero independent coverage; the badge wall is self-curated, not adoption.
  • Exact-version pins across layers (the CLI admits exactly one Ledger version) make upgrades an all-or-nothing dance.
  • The ecosystem is three repositories at 65, 1, and 2 stars: every layer is one maintainer’s bet.
  • 0.2.x with breaking changelog entries (autonomous run/resume refused without dispatch, watch demoted to read-only).
  • The company behind it (JUNO AI INC.) has no public funding record this note could find.

Pricing
#

Free and open source: the CLI is MIT, and the Ledger and Benchmark are MIT as well. No hosted tier and no paid plan found (as of 2026-10-21).

Compared to
#

  • Crewplane: both make the workflow the durable artifact; Crewplane renders each stage to disk as Markdown, while YYLO enforces a task lifecycle with validation and merge protection.
  • Worktrunk: owns the worktree lifecycle with hooks; YYLO adds task state, evidence, and protected merges on top of the worktree it creates per task.
  • The Perfect Orchestrator: the adversarial answer to worker overconfidence; YYLO’s verification is checklist- and evidence-based rather than worker-versus-worker.

Bottom line
#

Recommended for solo engineers who want bounded, receipt-backed agent tasks with protected merges and a Git-native paper trail. Not for teams wanting a GUI, and not without tolerance for 0.2.x churn and exact-pin upgrades.

Changes
#

  • 2026-10-21 - Created.

See also
#

  • Orchestration Feature Matrix - the category comparison this note joins
  • Crewplane - the Markdown-workflow runner in the same process-as-artifact seat
  • Worktrunk - the worktree lifecycle YYLO builds its task isolation on
  • The Perfect Orchestrator - the adversarial alternative to evidence-based verification
  • Bernstein - the deterministic-scheduler governance layer with the same trust-the-artifact instinct

References
#