Learn

Talk: Playbooks vs Existing Agentic Approaches

01 / Playbooks

What is a Playbook?

A Playbook is a preconfigured, ready-to-go team of agents that guarantees a specific process is followed as the agents work.

With many AI approaches (like /goal or dynamic workflows), you’re trusting an agent or harness to come up with a good process every time. The problem with this is an unsolvable alignment problem. Models are known to cheat, act lazily, become less coherent as context increases, and hallucinate. This is especially true when you use models that you didn't train yourself on new work that has never been done before. Without a strong verifier, outcome driven approaches are too risky to run. In practice this leads to wasted developer time, wasted tokens, increased rate of bugs, and generally more mistakes.

Playbooks solve the operational alignment problem by guaranteeing that work is done in a specific way by ensuring that each agent has a single responsibility, and agents hand off to one another correctly. With a Playbook, you choose the process ahead of time.

But a Playbook is not a fixed set of steps. Agents are spun up dynamically as the work reveals itself. Playbooks are more akin to policies produced by reinforcement learning. In RL, a policy specifies what action to take in each state; a Playbook specifies which agents should run given the current state of a task.

This is useful because when working on new things, you cannot know ahead of time what the task state will be (i.e. most software development is a discovery process). So much knowledge work, including software development, involves learning and discovering as you go. That makes the process incredibly important. You want agents following a process a domain expert has approved, one you can inspect, understand, and steer. Alinery is a Playbook IDE to make creating, inspecting, and steering agents effortless.

02 / Tasks

Orient by Task, not by Agent

We need a new way to orient ourselves when multiple agents work on a task. The old way of managing agents directly requires too much cognitive overload when multiple agent sessions are needed to get work done. Many existing apps assume 1 agent to 1 task and vice versa. Or they orient around projects. What's missing is the ability to orient at the task level.

Alinery is task-native by default. Each task gets its own context. Any number of agents can be spun up to operate within that context. And since so much of our work is discovery-based, we need a first-class way to explore rabbit-holes.

03 / Sub-Tasks

Introducing Sub-Tasks

Sub-tasks are a way to fork task level context. Sometimes building a new feature introduces a new bug and you have to go off and fix that, before you can merge the whole thing. But then you aren't sure of the right fix and you don't want to muck up all the work and context you've done at the top-level task. By creating a sub-task, you can explore based on the current work you've done, and easily discard it if it doesn't work out, or keep it and merge it back into the original task if it's keepable work.

Sub-tasks aren't just for humans; they make it easier for agents to orient. Every sub-task has a pointer to the original task context in case agents need it, but by default they know what the focus of the sub-task is, and are less likely to get confused with a massive amount of context at the higher task.

Sub-tasks can be infinitely forked supporting endless rabbit-holes. With Alinery, you never again have to worry about interrupting good progress when a hard problem arises; just spawn a sub-task.

04 / MCP

Manage your tasks through MCP.

Alinery is MCP-driven. Chat with an external agent connected through MCP to manage tasks, check in on progress, and make changes across multiple tasks at once.

When you need to get hands on or make a decision, jump into the UI. You can work through an agent or work directly with the task.

05 / Models & harnesses

Your models. Your harnesses.

Choose the models and agent harnesses you want to work with. By default Alinery supports and ships with the open-source harness OMP. Different work calls for different tools, and Alinery is built around that freedom.

You shouldn't have to organize all your work around a single model. Keep your tasks and Playbooks in Alinery, and choose the agents that do the work.

By default, Playbooks can specify which model should be used for each agent step.

06 / Playbook library

Turn your expertise into a Playbook.

Start with a Playbook from the library, adapt one to your work, or create your own. Define how the agents collaborate and where you step in. Share your Playbook so others can put your approach to work.