Inside Rezee - agents as teammates, not features (AI & Agents)
The thinking behind Rezee's AI & Agents layer - agents as workspace members with scoped access and full attribution, working assigned issues, CI failures, review, and doc drafts, with humans holding the gates.
Jul 31, 2026 · 3 min read · Kash Gohil
Rezee is one workspace for everything a team ships, organized as five layers. This post covers the newest and most consequential: AI & Agents. Its conviction is architectural: an agent is not a feature you sprinkle onto tools - it's a teammate you admit into a workspace, with an identity, scoped permissions, and full accountability. Platforms that treat AI as autocomplete-in-more-places are, we think, building for the previous era.
Why "teammate" and not "assistant"?
An assistant helps you while you work; a teammate works while you don't watch. That second thing - real delegation - is where the productivity step-change lives, and it has requirements assistants don't: the agent needs context (the spec, the issue, the conventions, the discussion) and the team needs control (what can it touch, who approves, what's the audit trail). Both are workspace properties, not model properties - which is why we argue the platform's shape decides what you can safely delegate.
What's actually in the layer?
Agents as workspace members. Each agent has its own identity - not borrowed credentials - with every action attributed and revocable in one step. The non-negotiable foundation of safe delegation.
Assign an issue, and it starts working. Assigning an issue to an agent hands it the title, the body, and the workspace - it investigates, comments its analysis, and adjusts status or labels where warranted. The highest-volume, lowest-risk delegation there is, and legible because issue structure makes it so.
First-pass review. An agent reads the diff, comments where something looks wrong, and records an approve-or-request-changes verdict under its own name - and can never cast one on a PR it opened, because agents are users and the author rule is the rule. The AI-then-human pattern that protects reviewer attention for intent and design.
CI failures, diagnosed. Set an agent watching CI and a failed run spawns a task: read the logs, work out what broke - flaky test, missing import, broken lockfile - and post the diagnosis and a suggested fix on the PR. It explains the failure; it doesn't push a fix. "Whenever someone notices" becomes minutes.
Doc drafts that track code. Set an agent watching pushes and every commit spawns a task to read the diff and draft the doc update it implies - the post-ship writing humans famously don't do. What it writes is an ordinary doc, for anyone to edit.
Human approval on anything irreversible. Merging a pull request is gated, always: the agent pauses itself at the gate rather than calling the tool, and its token is deleted while it waits. The agent proposes; a person disposes. "Humans keep the final say" is a permission setting in Rezee, not a policy memo. For readers who want the mechanism rather than the promise, the engineering posts go deep on both halves: how agent identity works at the schema level and how the approval gate works as a state machine.
Why does this need the other four layers?
Because context is the fuel. An agent working an issue reads the spec behind it; one diagnosing a pipeline sees the PR and its review; one drafting docs knows what actually merged; one reporting progress does it in the thread where the team already is. Bolted onto five separate tools, that context arrives as fragments through five API keyholes - the structural argument that runs through everything we've written about the agent era.
FAQ
What is Rezee's AI & Agents layer?
The delegation layer of the Rezee workspace: AI agents as first-class workspace members with their own identities and scoped permissions, working assigned issues, CI failures, code review, and documentation - with human approval required on irreversible actions.
Which AI models power Rezee agents?
Rezee is the workspace agents operate in - identity, permissions, context, and audit - and is designed to work with the leading agent models rather than compete with them. The tooling landscape moves fast; the workspace contract doesn't.
Do agents cost a seat?
No - bots and service accounts are free on every plan (pricing). You pay for humans; the agents work for free, which is rather the point.
Can I turn agents off entirely?
Yes - agents only exist in your workspace if you add them, their scopes are explicit, and revocation is one step. Teams adopt the layer in stages, usually starting with a single assigned issue.