---
title: "Narada AI: the enterprise agent behind WebArena's 86.1%"
description: "Narada AI builds enterprise automation agents and now tops WebArena at 86.1%. We have not tested it: what is public, and what the score does not say."
url: https://personalagentbench.com/narada-ai/
last_updated: 2026-10-19
---

# Narada AI, the company now at the top of WebArena

Narada AI is a Berkeley-born startup that sells agents to automate office work inside enterprise software. On 16 October 2026 it announced that its agents complete 86.1% of the tasks in [WebArena](https://personalagentbench.com/benchmarks/webarena/), the highest number on that benchmark's leaderboard and above the 78.24% human figure the benchmark reports. We have not tested Narada, it is not in our exam, and it is not a personal agent in the sense the rest of this site uses. This page sets out what is public about the company, who said each part, and what the 86.1% does and does not establish.

- Narada AI calls its product "agentic process automation": agents that click and type through enterprise apps such as ServiceNow, SAP Concur and Salesforce, sold mainly on custom enterprise contracts.
- The 86.1% comes from two products working together. Narada Code builds a workflow, and Narada's Operator agent runs it. Narada says so itself: "These are results for the combined workflow-development and execution configurations."
- The announcing post names no underlying model, run count or evaluation protocol, and we found no paper, released trajectories or third-party rerun.
- Narada's earlier WebArena claim, 64.2% for Operator alone in October 2025, was also self-reported.

## What Narada AI is, in its own words

Narada's [home page](https://narada.ai/) describes a platform for "agentic process automation" and positions it against older robotic process automation: "Moving beyond RPA, Narada was developed by AI researchers from Berkeley and Stanford." It says the agents work across "desktop apps, web, and Citrix," and that the automation is "self-healing" when an interface changes. The page is written for companies. Its sections are back-office automation, front-line automation and customer testimonials, and the main call to action is a demo booking.

The [developer documentation](https://docs.narada.ai/documentation/introduction) fills in the shape. You can drive the agents through a Python SDK or a REST API, install a Chrome extension, build agents visually in an Agent Studio, and run them in cloud browser sessions with "no local Chrome needed." It lists MCP servers and the A2A protocol for connecting other tools, and plugins for Gmail, Calendar, Slack and Zoom. It also cites the team's research background, the LLMCompiler and Plan-and-Act papers.

The [Chrome Web Store listing](https://chromewebstore.google.com/detail/narada/cmgiedoanffjpkfakjjbfiiaccklfdfo), read on 19 October 2026, is titled "Narada AI Assistant: Computer Use for Enterprise Tools." It showed 1,000 users, a 5.0 rating from 12 reviews, and version 1.0.263, updated 15 October 2026. Its examples are processing invoices, updating records and managing approvals.

## Who is behind Narada AI

The [leadership page](https://narada.ai/leadership) names three founders. David Park is chief executive; the page says he co-founded Coverity, the code-analysis company, out of a Stanford computer science PhD. Amir Gholami is chief technology officer and is described as a research scientist at BAIR and the Sky lab, both at UC Berkeley. Kurt Keutzer, chief science officer, is an AI professor at UC Berkeley.

TechCrunch [covered the launch](https://techcrunch.com/2024/10/29/narada-ais-enterprise-agent-will-use-workplace-tools-for-you) on 29 October 2024, when Narada was one of the Startup Battlefield 20 at TechCrunch Disrupt. The article connects the product to the LLMCompiler paper Keutzer and Gholami co-wrote, reports that "the startup already has a Fortune 500 company using its agent" without naming it, and says the company had "raised a few million dollars from a few advisers" and was talking to venture investors.

Park told the story again on TechCrunch's [Build Mode podcast](https://techcrunch.com/video/building-a-lean-team-before-raising-big-with-narada-ceo-david-park/) in March 2026. The episode page says Narada "spent a year making nearly 1,000 customer calls before raising institutional capital," and credits it with 99.99% reliability in production, a figure the page does not source further.

We found no funding announcement from Narada itself. Two deal-tracking sites that rank for the company's name list a round of about $8 million, but neither cites a filing or a press release we could open, so we do not treat the figure as confirmed.

## What Narada AI looked like before the enterprise pivot

The 2024 TechCrunch piece describes something closer to the products on this site than Narada's current marketing does. "The assistant sits in a separate chat window in your browser," it says, and the reporter used it to draft an email declining an invitation, find a restaurant, create a calendar invite and email a friend. It used APIs where it could and clicked through web pages where it could not.

Even then, the scope was work software. TechCrunch wrote that Narada was "solely focused on enterprise applications, instead of a general purpose agent for any website," and its reporter's attempts on LinkedIn and Facebook produced errors. Two years later, the site, the docs and the extension listing all describe a tool a company deploys for its staff, with an on-premises option, rather than an assistant an individual signs up for and points at their own life.

That is why Narada is not on our [landscape of personal agents](https://personalagentbench.com/landscape/). It does the kind of work our errands test, filling forms and moving records between apps, but its buyer and its setting are different.

## Narada AI pricing and where it runs

The pricing section on Narada's home page, read on 19 October 2026, has two tiers and no published prices.

| Tier | Price | What the page lists |
| --- | --- | --- |
| Free | $0 | "Get hands-on with core features," 400 Narada Credits, limited API access |
| Enterprise | Custom | Custom credits, API access, priority support, on-prem deployment, BAA, DPA, GDPR and HIPAA terms, contract redlining, a dedicated account manager, use-case tuning |

The free tier's button leads to the Chrome extension. We could not find what a Narada Credit buys, or the price of additional credits. The Narada Code announcement adds one relative figure: OperatorMini "consumes 6-8x fewer credits than Operator."

## Narada Code and OperatorMini, the two products behind the score

The [16 October 2026 post](https://narada.ai/blog/introducing-narada-code-operatormini) introduces two products.

Narada Code is "a computer-use agent harness that builds, tests, and improves enterprise workflows." In the post's description, it investigates the applications involved, drafts a workflow, tests it in a sandbox, and repairs it when runs fail. It is meant to work alongside existing coding harnesses, and the post names OpenAI Codex and Claude Code. "Narada Code is now available to our enterprise customers," the post says, with no self-serve route; the call to action is to book a call.

OperatorMini is "a lightweight execution agent designed to reduce the cost of running them," meaning the workflows Narada Code produces. It needs more explicit step-by-step instructions than the full Operator. The post suggests OperatorMini for well-defined, repetitive work and the full Operator for tasks that need more judgement.

So there are three named pieces: Operator, the general execution agent first announced in 2025; OperatorMini, the cheaper executor; and Narada Code, which writes the workflows the other two follow. The leaderboard row is labelled "Narada AI Code," but the post is clear that the score belongs to the combination.

## What Narada claims on WebArena

The post's numbers, as Narada states them:

| Configuration | WebArena success | Source |
| --- | --- | --- |
| Operator, with workflows developed by Narada Code | 86.1% | Narada, 16 October 2026 |
| OperatorMini, with workflows developed by Narada Code | 83.5% | Narada, 16 October 2026 |
| Previous best on the leaderboard, as Narada describes it | 74.3% | Narada, 16 October 2026 |
| Operator alone | 64.2% (64.16% on the blog) | Narada, October 2025 |
| Human baseline | 78.24% | WebArena paper, 2023 |

On the multi-site tasks, which ask an agent to work across more than one of the benchmark's websites, the post gives 85.4% for Operator and 81.3% for OperatorMini, against 54.2% for WebTactix. It compares with Claude Code plus GBOX MCP, IBM CUGA and OpenAI's Operator in charts.

The [leaderboard sheet](https://docs.google.com/spreadsheets/d/1M801lEpBbKSNwP-vDBkC_pF7LdyGU1f_ufZb_NWNBZQ/edit) lists the row as "Narada AI Code," 86.1%, dated September 2026, with Narada's own team as the result source and the "Open?" column marked no. An older "Narada AI" row at 64.2% is marked self-reported.

The earlier claim followed the same pattern. In [October 2025](https://x.com/Narada_AI/status/1976817576916459864) Narada announced 64.2% for Operator and called it a new state of the art. Its [blog post](https://narada.ai/blog/narada-ai-web-agent-operator) gave 64.16% and also reported 97.45% on WebVoyager, against 87% for OpenAI Operator and 89.1% for Browser Use. Neither post describes the evaluation protocol or the model underneath.

## What the WebArena number does not tell you

Four things are missing, and each matters for how far the 86.1% travels.

The first is the setup. "With workflows developed by Narada Code" can mean very different experiments. A workflow could be built once per website, once per task template, or once per task, and it could be built with or without access to the benchmark's tasks and its grader. WebArena has 812 tasks generated from about 170 templates, all run on self-hosted copies of a small set of websites, so a workflow tuned on those sites is tuned on the test environment. The post does not say which of these Narada did, or whether any tasks were held out. Neither answer would make the result wrong; they would make it a different kind of result.

The second is the model. Neither the 2025 nor the 2026 post names the model or models behind Operator, OperatorMini or Narada Code.

The third is replication. We found no paper, no released trajectories and no independent rerun as of 19 October 2026. The leaderboard row is self-reported. That is common on this leaderboard, but it means the only score above the human figure is also one a reader cannot check.

The fourth is the grader. A Berkeley team [showed in April 2026](https://rdi.berkeley.edu/blog/trustworthy-benchmarks-cont/) that the original WebArena grader could be gamed: an agent could read the reference answers out of the task configuration and reach roughly 100% without doing the tasks. We have no reason to think Narada did anything of the kind. It does mean a high score on the original sheet is worth less without a trace showing the work. ServiceNow's [WebArena-Verified](https://github.com/ServiceNow/webarena-verified) re-release fixes the grader, and we found no Narada result on it.

The human comparison needs its own caveat. The 78.24% came from five graduate students, each given one sampled task per template, not all 812 tasks. Our [WebArena page](https://personalagentbench.com/benchmarks/webarena/) explains why an agent number above it is not the same as beating people at WebArena.

## Is Narada AI a personal agent?

Not by the test this site applies. The products we take apart are ones a person signs up for and lets act in their own email, accounts and errands: [Muse](https://personalagentbench.com/agents/muse/), [Instinct](https://personalagentbench.com/agents/instinct/), [Pine](https://personalagentbench.com/agents/pine/) and [Town](https://personalagentbench.com/agents/town/). Narada sells to companies, runs inside company systems, and offers on-premises deployment and compliance paperwork as part of the product. Its free tier is a trial of the enterprise tool rather than a consumer plan.

What connects Narada to our subject is the method. Its agents operate software through the screen when there is no API, which is the same capability our [browser tasks](https://personalagentbench.com/how-the-exam-works/browser-tasks/) test in consumer settings. A benchmark result in one setting does not transfer to the other. A workflow tuned for a known set of enterprise screens is a different job from an agent meeting a restaurant's booking page for the first time, which is part of why we report [measured completion with intervals](https://personalagentbench.com/how-scoring-works/) on named errands rather than a single leaderboard figure.

## FAQ

**What is Narada AI?** An enterprise software company that sells agents for "agentic process automation," automating work in tools such as ServiceNow, SAP Concur, Salesforce, Slack and Outlook through APIs and through the screen. Its founders are David Park, Amir Gholami and Kurt Keutzer, with roots in UC Berkeley research.

**What is Narada Code?** Narada describes it as "a computer-use agent harness that builds, tests, and improves enterprise workflows." It was announced on 16 October 2026 and is available to Narada's enterprise customers.

**Did Narada AI really score 86.1% on WebArena?** That is Narada's own reported figure, for its Operator agent running workflows that Narada Code developed. The leaderboard lists it as self-reported, and we found no paper, released trajectories or independent rerun as of 19 October 2026.

**Does 86.1% mean Narada beats humans at web tasks?** Not on this evidence. The 78.24% human figure was measured on about 170 sampled tasks, not the 812 agents run, and the setup behind the 86.1% is not published.

**How much does Narada AI cost?** Narada's site lists a free tier with 400 Narada Credits and a custom-priced Enterprise plan. It does not publish what a credit costs.

**Can I use Narada as a personal assistant?** You can install the Chrome extension and try the free tier, but the product is built and sold for enterprise workflows. We have not tested it. For personal agents we have opened, start with [the teardowns](https://personalagentbench.com/agents/).

**Is Narada AI related to narad.io or Narada Systems?** We found nothing linking them. Several unrelated companies use the Narada name, including narad.io, a third-party risk vendor; narada.systems; and NaradaCode, an IT consultancy.

---

Published by Personal Agent Bench (https://personalagentbench.com), run by the Zentor team. Zentor builds agent software of its own, so its products are not on the scorecard and will not be added.