Warp started out as a faster, prettier terminal. It used blocks instead of raw text, a modern input field, GPU rendering, and a UI that didn’t feel stuck in the 80s. But that version of Warp is basically gone now. What you’re looking at today is something else entirely.

After spending time with its new agentic coding mode, Warp is no longer trying to improve the terminal. It’s trying to replace how you work inside it. The terminal is still there, but it’s no longer the center of gravity. The real product is an agent workspace that can read your codebase, run commands, propose changes, and ask for approval along the way. Whether that’s better depends on how much control you’re willing to give up.

Warp is no longer what it used to be

It's now an "agentic development environment"

The easiest way to understand Warp today is to stop thinking of it as a terminal. It’s closer to a development environment where agents sit at the center and everything else exists to support them.

You still get the familiar terminal surface, but even that has been reworked. Commands and outputs are grouped into blocks instead of a continuous scroll. It feels less like typing into a terminal and more like interacting with a UI that just happens to execute shell commands.

Then there’s the shift in how work is organized. Warp introduces a separate “agent mode” where you can have multi-step conversations with an agent that can break tasks down, create plans, and execute them with checkpoints. You can run multiple of these in parallel, which is clearly where Warp thinks development is headed.

So Warp now effectively includes a lightweight code editor, a file tree, and a diff review panel. When an agent makes changes, you can review them right there, leave comments, and ask the agent to revise.

There’s also a knowledge layer sitting underneath all of this. Warp indexes your codebase into embeddings, so the agent can retrieve context without guessing. It introduces “rules” and “skills” that can be stored in your repo or in Warp Drive, which acts as a shared configuration layer for teams. Rules define how the agent should behave, while skills act like reusable workflows.

And then there’s Oz, the orchestration layer that changes what Warp is capable of. This is what allows agents to run in the cloud, be triggered by external tools, and operate outside your local machine.

Warp’s agentic coding is something you can actually use

It supports OpenAI, Anthropic, Google, and other hosted options

Warp’s agentic coding mode is a collection of systems that allow an LLM-guided process to operate inside a real repository and terminal. At the model layer, Warp supports multiple providers with explicit model identifiers. You can choose specific models or let Warp handle routing automatically. It supports OpenAI, Anthropic, Google, and other hosted options, with fallback mechanisms if a selected model is unavailable.

The tool splits execution between local and cloud environments. Local agents run directly inside the Warp app and can interact with your shell and workspace, with permissions and review gates in place. Cloud agents run remotely in configured environments, either triggered manually or through integrations and schedules. These runs are logged with full transcripts for auditability.

One of the more powerful capabilities I have used in Warp is Full Terminal Use, which operates at the PTY level. This allows the agent to interact with live terminal sessions, including database shells, debuggers, and long-running processes. It can read the terminal buffer, respond to prompts, and continue execution within the same session, which is what enables the “agent drives the terminal” experience.

For larger tasks, Warp leans heavily on planning. Prompts can be converted into structured, editable plans with version history and selective execution. It can also generate task lists automatically for complex requests, giving you visibility into progress and intermediate steps.

Warp handles code changes through a diff-first workflow. As the agent generates code, you can review changes in a dedicated panel, leave inline comments, and provide batch feedback. The agent then applies revisions in subsequent iterations.

Warp also builds an embeddings-based index of your codebase, referred to as Codebase Context. This allows the agent to retrieve relevant code without relying entirely on ad-hoc search. While Warp states that raw source code is not stored, it is transmitted during indexing to generate embeddings, which are retained. In practice, this means your code is processed server-side, even if it is not stored as raw files.

On the extensibility side, Warp supports MCP servers, which act as plugins for adding custom tools and data sources. These can connect over multiple protocols and include granular permission controls, allowing you to define whether access is always allowed, denied, or decided by the agent.