Runway Agent 2.0 is a creative product for marketers rather than a new video model. It sits above multiple models, asks questions about a product and audience, analyzes supplied campaign material, and generates the concepts and assets needed for the next marketing iteration.

Runway introduced Agent 2.0 on June 25, 2026 and made it available to all users. The product is organized as one conversation that can begin with a rough idea, an underperforming ad, a launch brief, a target audience, or an existing set of creative assets. The agent is intended to move between analysis and creation instead of requiring the marketer to choose every model and tool manually.

That makes Agent 2.0 closer to a specialized creative workspace than a universal autonomous worker. Its launch capabilities center on marketing strategy, content variants, and generation inside Runway. The company’s more ambitious vision—directly connecting campaign platforms, learning continuously from performance, and generating improvements automatically—is described as what comes next.

From a brief to a campaign angle

Campaign work rarely fails because no one can generate an image. The harder problem is deciding what to say, to whom, in which format, and how each asset supports a testable idea. Runway positions Agent 2.0 as a partner that asks questions and develops positioning before producing media.

A founder can describe a product and audience, then work through possible hooks. A product marketer can explore messages for a new launch. A brand team can develop seasonal or localized variants. The conversation retains the goal while the user moves from concept to generated deliverables.

The quality of the brief still determines the useful search space. Provide approved product facts, audience evidence, brand rules, prohibited claims, required formats, and a definition of success. An agent cannot infer regulatory limits or genuine differentiation from a product photo alone.

Separate facts from hypotheses. “Customers save 20 percent” needs evidence; “test a convenience-focused hook” is a creative proposal. Generated copy should not turn an experiment into a factual claim.

Analyzing what is already working

Agent 2.0 accepts creative and performance data from campaigns on Meta, YouTube, TikTok, or Google. Runway says it can analyze what is working and create the next set of ads to test. That can compress a workflow that normally spans dashboards, spreadsheets, creative review, and several generation tools.

Performance data needs context. A high click-through rate can accompany poor conversion. A winning ad may have benefited from audience targeting, placement, seasonality, or spend rather than the visual hook alone. Give the agent metrics with definitions, date ranges, sample sizes, and funnel outcomes.

Use its analysis to form testable recommendations, not causal certainty. A useful output says, “These three high-performing ads share a direct demonstration in the first two seconds; test that structure against the current control.” A weak output says, “Blue backgrounds cause purchases.”

Protect campaign data as business-sensitive information. Remove customer-level data, follow platform terms, and verify Runway’s current retention and workspace controls before uploading exports.

Keep a human-readable analysis record outside the conversation as well. It should identify the source report, metric window, chosen hypothesis, assets generated, and final decision. That record lets another reviewer challenge the reasoning without reconstructing an entire agent session and prevents a later campaign from treating an old correlation as a permanent rule.

Generating variants for channels and markets

Agent can produce video ads, product videos, social posts, image ads, commercials, and campaign sets. Runway highlights automatic sizing for 9:16 Reels and Stories, 16:9 YouTube, and 1:1 feed placements. It can also localize copy and swap visuals for different markets.

Format adaptation involves more than cropping. A vertical placement may need a tighter subject, larger text, different pacing, and safe areas for interface overlays. A localized campaign may need different claims, products, prices, cultural references, and disclosure language. Review each variant in its destination context.

Runway’s product page says Agent can choose among models such as Gen-4.5, Aleph 2.0, Seedance 2.0, Nano Banana Pro, and Kling 3.0 Pro, or use a model requested by the user. Automatic selection reduces setup, but it can make reproducibility harder. Record the model and settings behind every approved asset where the product exposes them.

Iteration inside one conversation

The user can request changes such as a new voiceover or stronger opening, and Agent revises the work. This conversational loop is useful because feedback remains connected to the campaign intent. It is also a potential cost trap: “try another” can generate many assets without improving the hypothesis.

Structure each round:

  1. state what failed the review;
  2. change one meaningful variable where possible;
  3. preserve approved brand elements;
  4. define the new acceptance criterion;
  5. record the candidate and generation cost;
  6. stop after a planned number of variants.

Creative testing needs diversity, but uncontrolled generation produces a library no one can evaluate. Name each variant by hypothesis rather than by aesthetic adjective.

Availability and cost

Runway says Agent 2.0 is available for all users and across Runway plans. The product uses generation capabilities tied to the account, so cost depends on plan, selected model, resolution, length, and number of variants. The launch post does not establish one fixed price for an Agent conversation.

Budget by campaign output: accepted concepts, usable assets, and variants sent to test. Track generations that were rejected, revisions, and localization overhead. A low cost per generated video is not valuable if the campaign team cannot identify which version is approved.

Users should also separate current capabilities from the roadmap. Runway says it wants Agent to connect directly to marketing platforms, learn from performance data, and keep campaigns improving. Until a specific connection and action is documented as available, assume the user supplies data and controls publication.

Governance for a marketing agent

Campaign generation can introduce false product claims, unlicensed references, inconsistent logos, or synthetic people who appear to endorse a product. Establish review owners for factual, brand, legal, and media-rights checks.

Do not let historical performance override brand or safety policy. An agent may observe that aggressive claims get clicks; that does not authorize them. Keep prohibited content, required substantiation, and market-specific rules in the brief and final checklist.

For high-volume variants, automate mechanical checks such as dimensions, duration, file format, and naming. Use humans for meaning, identity, tone, and rights.

Agent 2.0 versus a general tool orchestrator

Runway Agent is specialized around creative analysis and generation within Runway. A general agent such as Codex or Claude can instead coordinate separate services and deterministic operations through MCP. Neither architecture is universally better.

Use Runway Agent when the main job is developing campaign ideas, selecting generation models, creating assets, and iterating creatively in one product. Use a general orchestrator when the workflow spans repositories, documents, file systems, approval logic, or post-production tools from several vendors.

Medux is one external media-processing service that Codex or Claude can use after separate configuration; it is not natively connected to Runway Agent. Its Codex MCP video merge tutorial demonstrates an ordered, asynchronous merge of approved clips. Its Claude MCP logo tutorial applies a supplied brand mark to a video.

A hybrid campaign workflow could be:

  1. use Agent 2.0 to analyze supplied aggregate metrics and propose testable angles;
  2. generate and review visual candidates in Runway;
  3. export only approved clips with model and rights records;
  4. let Codex or Claude assemble clips in an approved order;
  5. add the exact current logo asset as a deterministic step;
  6. verify aspect ratio, duration, logo position, and filename;
  7. require platform-specific human approval before upload.

This keeps creative exploration flexible while making brand finishing and delivery auditable. Runway Agent 2.0’s promise is not that marketers can disappear. It is that the gap between campaign evidence, a new hypothesis, and the assets needed to test it can become much smaller.