Implementation

Dig into the tech

How the PPO policy picks plays, the three-layer ledger, cross-framework review, the autonomy posture, and how to run it.

Open source Read the source on GitHub →

What's different

A learned manager, not a smarter worker

Five design choices that separate Swink AgentShore™ from a prompt-engineered loop or a hand-authored DAG.

Decisions are learned, not scripted A PPO actor-critic chooses among 19 active plays (22 action slots) each tick. Cold-start logits bias toward issue pickup, code review, QA, and merge. The policy improves with every session. How the policy works →

BEADS is the canonical graph Epics, stories, and tasks live in a beads-native graph. GitHub is the human surface. Swink AgentShore™'s SQLite holds the RL state. Alignment is tracked as a live per-tick delta. The three-layer model →

Claude doesn't grade Claude Code review and QA run on a different framework than the author. Confirmation bias compounds when an agent reviews itself. The failure looks like success right up until merge. Cross-framework review →

YOLO by design, with hard gates The default control loop is autonomous. Issues get created, edited, labeled, and merged without human approval. Hard gates cover anti-confirmation, merge readiness, budget and time exhaustion, and audit. Autonomy & gates →

Nothing inherited at dispatch Every spawned subprocess gets a six-piece envelope assembled per play: rendered SKILL.md, beads task graph, GitHub mirror, identity overlay, working directory, and limits. Nothing leaks in from the calling shell. The dispatch envelope →

Multiple harnesses, effort that tracks the work Codex, Claude Code, and Antigravity agents run against one backlog at once. Concurrency and level of effort ebb and flow with the real state of the graph. The policy pushes when work is pickable and backs off when it's blocked. Concurrency over a run →

How you use it

Desktop app first. CLI when you want it.

Most of the time you'll launch Swink AgentShore™ from the dock. The desktop app bundles the session, the pixel-art dashboard, and the project graph into one window. You watch the policy work in real time. The CLI is the same tool, headless, for CI and scripting.

Default surface

Swink AgentShore™.app

Tauri-based. macOS (Apple Silicon) and Windows available; Linux planned. Bundles the bd sidecar so there's nothing to install separately.

Live pixel-art dashboard of the running session
Session start, stop, and status
Configs for agent tiers, budget, and identity

Inside the Dashboard →

For headless & CI

agentshore CLI

The full orchestrator surface for CI, scripting, and power users, plus an interactive TUI for watching a session from the terminal.

# initialize a new project
$ agentshore init

# start a session
$ agentshore start

# open the dashboard against a running session
$ agentshore dashboard

Same window, from the terminal →

The data model

Three sources of truth

Different three from the framework layers above. This is where project state actually lives. Each source owns one thing; the boundary is the interesting part.

BEADS

The project graph Epics, stories, tasks. The canonical structure of the work. Swink AgentShore™ reads it; humans don't edit it directly.

GitHub

The human surface Issues, PRs, comments. The conversation layer where humans inspect and steer. Mirrored to and from BEADS.

SQLite

The RL state Session-scoped policy state, experience replay, audit records. Local. Inspectable. Versioned per release.