# How Obss works - Obss

Get started
How Obss works
The Obss loop, where files, machines, and tools attach, and the main terms you’ll see across the product.
Obss has four main parts: tasks (the work), treatments (the tools an agent gets), agents (such as Claude Code or Codex), and the runs they produce. When you know how they fit together, you know exactly what a result compares.
The loop
Stage
What you do
Start here
Define tasks
Write the work: instructions, a rubric, files, and a machine.
Tasks and rubrics
Build treatments
Choose the tools agents get.
Raw Agent
gives none and is the usual baseline.
Treatments (Toolsets)
Run experiments
Pick the tasks, treatments, agents, and trials to run.
Launch an experiment
Read results
Check
Matrix
,
Comparisons
,
Frictions
, and each run’s details.
Read results
Improve
Fix what the frictions point to, then launch again or on a
schedule
.
Find and fix frictions
You can do all of this from the dashboard,
Ask Obss
, the
Obss Plugin
in your AI assistant, or the
CLI
.
Where files, machines, and tools attach
Each part of a run has one home. Files and a machine belong to a task, tools belong to a treatment, and you choose agents when you launch.
You create files under
Manage → Files
, machines under
Manage → Machines (Environments)
, and tools under
Manage → Tools
. Agents can be built-in or
custom
(private beta). Learn more in
Machines and files
,
Tools
, and
Agents
.
Because each part has one home, a comparison between two treatments changes only the tools, and a comparison between two agents changes only the agent. An experiment runs every combination of task, treatment, agent, and trial, and each combination is one run. To learn how runs are counted and why the comparison is fair, read
Experiments
.
To measure a tool’s impact, choose a
baseline
at launch and compare each treatment’s pass rate with it on
Matrix
. The difference is the treatment’s lift.
Main terms
Term
Meaning
Task
The work an agent attempts: instructions, a rubric, files, and a machine
Rubric
The criteria that Evals use to mark a run Pass or Fail
Tools
Skills, MCP servers, CLIs, and SDKs that you give an agent
Treatment (toolset)
The set of tools an agent gets in a run.
Raw Agent
gives none.
Agent
The AI agent that does the work, such as Claude Code or Codex, with a model and an effort level
Experiment
One launch of tasks, treatments, agents, and trials
Run
One agent’s attempt at one task with one treatment, and the unit you pay for
Trial
A repeat of the same inputs. Each trial is one run.
Lift
The difference in pass rate between a treatment and the baseline
Friction
One problem that agents hit, tracked across runs and experiments
The
Glossary
defines every other term.
Next steps
Set up your organization
Finish the one-time setup checklist.
Tasks and rubrics
Write the work and the rubric that decides whether a run passes.
Treatments (Toolsets)
Build the sets of tools you want to compare.
Dashboard map
Find your way around the dashboard.
Quickstart
Dashboard map
Powered by
This documentation is built and hosted on Mintlify, a developer documentation platform