---
type: "post"
id: "framer-harness-2-applies-agent-edits-while-the-5223b7"
slug: "framer-harness-2-applies-agent-edits-while-the-5223b7"
title: "Framer Harness 2 applies agent edits while the model is still writing"
description: "Framer says it rebuilt the harness for its agent so a single program can read a document, make decisions, and apply changes while the model is still generating later steps."
retrieval_nugget: "Framer says it rebuilt the harness for its agent so a single program can read a document, make decisions, and apply changes while the model is still generating later steps. The company reports half as many tool steps and 39% fewer cached input tokens per message, while exposing the same capabilities to external agents through MCP."
published_at: "2026-11-04"
updated_at: "2026-11-07"
record_date: "2026-11-04"
date_kind: "published_at"
topics: ["mcp"]
entities: ["Framer","Framer Agent","Harness 2","MCP"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "same_event"
prior_refs: []
source_url: "https://www.framer.com/updates/harness-2"
source_urls: ["https://www.framer.com/updates/harness-2"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:framer-harness-2-applies-agent-edits-while-the-5223b7"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/framer-harness-2-applies-agent-edits-while-the-5223b7/","markdown":"https://newruntime.com/posts/framer-harness-2-applies-agent-edits-while-the-5223b7.md","json":"https://newruntime.com/posts/framer-harness-2-applies-agent-edits-while-the-5223b7.json"}
markdown_url: "https://newruntime.com/posts/framer-harness-2-applies-agent-edits-while-the-5223b7.md"
---

# Framer Harness 2 applies agent edits while the model is still writing

## Retrieval answer

Framer says it rebuilt the harness for its agent so a single program can read a document, make decisions, and apply changes while the model is still generating later steps. The company reports half as many tool steps and 39% fewer cached input tokens per message, while exposing the same capabilities to external agents through MCP.

## Summary
Framer says it rebuilt the harness for its agent so a single program can read a document, make decisions, and apply changes while the model is still generating later steps. The company reports half as many tool steps and 39% fewer cached input tokens per message, while exposing the same capabilities to external agents through MCP.

## What happened
Framer’s November 2, 2026 product update announces Harness 2, a rebuilt harness for the Framer Agent. The update says it can take on whole workflows in a single program. It reads a document, makes decisions, and applies changes while the model is still writing the next steps; Framer says users therefore see progress sooner. The update does not provide a detailed implementation account, but presents the change as a shift in how the agent executes work rather than simply a new model.

Framer reports two efficiency figures: Harness 2 uses 2x fewer tool steps and 39% fewer cached input tokens per message. The supplied update does not specify a benchmark protocol, baseline, or whether those figures represent averages over particular tasks, so they should be read as Framer’s claims rather than independently established performance results. Framer also says the capabilities are available to external agents through MCP, extending the harness’s actions beyond Framer’s own agent.

The update lists expanded operations across the product. For components, the agent can edit object and array fields, work with more types, and change animation settings. In the CMS it can configure sorting by multiple fields, status, and creation date. For forms, it can configure custom email recipients and webhook destinations. It can read and edit code, use exports at once, and add or update SVGs. It can create, switch, rename, join, leave, and delete branches. It can also add and customize more properties across project styles.

The update specifies role-based limits for CMS changes: content editors can edit and undo them, while viewers can ask but cannot edit. It does not state a separate price, a versioned rollout schedule, or broader access conditions. Framer links to a technical blog post for a closer look, but the supplied public product update is the basis for the capabilities and figures here.

## What the archive adds
The archive includes [Unreal Agent: Async-first agent harness](https://newruntime.com/posts/unreal-agent-async-first-agent-harness-analysis-78278b/), which treats long-running agent work as a runtime and persistence problem, and [Microsoft’s Copilot harness in Foundry Agent Service](https://newruntime.com/posts/introducing-the-microsoft-copilot-harness-in-foundry-agent-f2f750/). These are related context, not earlier coverage of this Framer release. Harness 2 is a new event: it adds concrete editing actions and reported execution-efficiency gains to the broader harness discussion.

## Trend signal
Framer’s update supports [Harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/): the harness governs workflow, tool actions, and external-agent access, while the announcement does not center on a model change. It is evidence about the product boundary and integration surface, though it does not establish that teams can freely swap models behind this particular agent.

## Why now
Framer describes the rebuild as a way to take on whole workflows and show progress sooner by applying changes while later steps are still being generated. The update gives no further explanation of the timing or an external event that prompted the release.

## Why it matters
For teams building AI-enabled products, the announcement is a reminder that perceived speed and useful capability depend on orchestration as well as model generation. Applying changes before a full sequence of steps is complete can make work visible earlier, while fewer tool steps and cached input tokens may reduce interaction overhead. Those benefits remain claims from Framer, not a general result proven across agents or workloads.

The breadth of operations also matters: an agent’s usefulness depends on the surface it can safely manipulate. Editing code, CMS settings, forms, styles, branches, and SVGs is more consequential than returning a recommendation in chat. Role distinctions—editors can change and undo CMS data, viewers can ask but not edit—show that access boundaries remain part of the workflow. External MCP availability could let other agents compose those capabilities, but the update does not explain authorization details or operational safeguards.

## Limits and uncertainty
The announcement does not disclose how Harness 2 is implemented, which models it uses, or how the single-program workflow handles failures, retries, conflicting edits, or partial completion. It gives no evaluation set, task mix, baseline, or measurement method for the 2x tool-step and 39% cached-token claims. The role limits are described only for CMS changes; the update does not establish permission behavior for every operation or explain MCP authentication. It also gives no pricing, rollout details, or evidence that the improvements generalize beyond Framer’s own product environment.

## Reader takeaway
Framer is presenting the harness—not just the model—as a place where workflow speed, product-specific actions, and agent access are designed together. The headline metrics are worth noting, but need benchmark context before they can guide comparisons. For builders, the concrete lesson is to define what an agent can do in the product and where permissions apply, then measure completed workflows rather than relying on model output alone.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Mcp](https://newruntime.com/topics/mcp.md): Explore the mcp topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)