Swink AgentShore
An experiment in autopilot for a multi-agent workforce.
What could a one-person scrum team look like with AI?
Swink AgentShore isn't trying to replace the Human Engineer. The experiment is whether one Human Engineer can ride point on a full delivery loop, with a Reinforcement Learning (RL) Agent sequencing the work and LLM Agents doing it. Not another coding agent. Long running orchestration. Less vibes, more ability to be away from the keyboard.
The 80/20 isn't a guarantee. It's the shape of the bet. The interesting question is which 80% the RL Agent can actually take, and which 20% it shouldn't even try.
Not every agent is an LLM
"AI agent" has narrowed to mean LLM. That's not wrong. It's just incomplete. Reinforcement learning agents are a different shape of intelligence: no context window, a partial view of an environment, and a choice of actions. They're built for the job LLM Agents aren't. Watching a system over time, learning which actions move it forward, and picking the next best move under uncertainty.
The muscle
Codes, reviews, debugs, calls tools. Anything you'd hand to a senior individual contributor with an open editor and a terminal.
- Sees
- A context window: the prompt and the history.
- In Swink AgentShore
- A mix of Claude, Codex, and Gemini, sized by the RL Agent to balance cost and throughput.
The manager
Watches the project state and picks the next play. Doesn't write code. Doesn't read your prompts. Learns from the trajectory of decisions it has already made.
- Sees
- A partial view of an environment and a choice of actions.
- In Swink AgentShore
- One small actor-critic network. Inspectable, replayable, tuned to this project.
RL isn't replacing LLM Agents here. It's coordinating them. Different tools doing what they're optimal for.
Four shores, four different jobs
Where labor lives has always been a tradeoff between cost, coverage, and collaboration. Each shore is an answer to a different mix, not a ranking. Different work, different shore. AgentShoring is a new option in the mix, not a replacement for the others.
Two terms in play. AgentShoring is the labor category (the new shore, where this row sits in the table below). Swink AgentShore is this product, one implementation of it.
| Shore | Strengths | Tradeoff | Best for |
|---|---|---|---|
| Onshore | Same-room collaboration, native cultural fit, real-time judgment | Cost | Tight in-person feedback loops |
| Nearshore | Same-day overlap, cultural proximity, easier travel | Smaller talent pools than offshore hubs | Same-day collaboration |
| Offshore | Scale, deep talent pools, 24/7 coverage when paired with onshore | Async coordination, more structured handoffs | Steady-state throughput |
| AgentShoring | Local execution, replayable decisions, transparent budgets | Bounded by training data; needs clear goal specification | Measurable, on-demand work |
AgentShoring sits beside the others, not above them. It optimizes for measurability and on-demand execution. The work happens locally, against a budget you set, with every play replayable from its checkpoint. That solves a different problem than offshore scale or nearshore overlap. Pick the shore that fits the work.
Swink AgentShore is the management tier. It watches the repo, the budget, and the open issues, then picks the next play and the LLM Agent that should run it.
What stays human
The shoring metaphor is honest about its limits. Three things stay onshore, forever:
- The problem statement. Someone has to write the issue, the PRD, the goal. Swink AgentShore reads them; it doesn't author them.
- The reward function. What counts as "shipped" is a human decision. The RL Agent learns to maximize a signal someone designed.
- The override. Humans can pause, redirect, or kill a session. Hard gates exist for things that must never happen, even on human override.
The honest negative
AgentShoring is not always the right decision. Agents can excel for a clear, well-structured PRD, a single repo, and an overnight run. There are varying levels to this, and vibe coding or a Ralph loop can work. Learned orchestration earns its complexity when the right next move isn't obvious enough to script, the budget matters, and the cost of merging the wrong thing is high. However, the thing to build and what good looks like needs to stay human.
Dig into the tech
How the RL Agent picks plays, the three-layer BEADS · GitHub · SQLite ledger, cross-framework review, the autonomy posture, and how you actually run it.
Open the tech overview →