Runway AI announced Agent Skills on July 2, 2026, describing campaign building, commercial creation, and ad localization from a simple command. That is enough to justify a controlled workflow test—but not enough to hand an agent a brand account and hope it makes every judgment correctly. The useful question is where delegation saves work without removing human responsibility.

Create Your AI Influencer

This guide converts the official Runway changelog into a bounded creator test. It uses one fictional adult persona, one owned product, locked inputs, and explicit approval gates. It does not claim that an agent can publish autonomously, guarantee legal compliance, or replace the people responsible for brand, rights, accuracy, and disclosure.

About this guide: APOB AI prepared this protocol from the first-party sources below, rechecked August 26, 2026. It contains no private benchmark data; teams must attach their own outputs, timings, failures, and decisions.

What Runway shipped on July 2

Start with a dated evidence box. Record the source, the exact named workflows, the account path tested, and anything the source does not establish. This prevents marketing language from becoming an undocumented operating permission.

Named skill examples

Runway’s July 2 changelog entry says Agent Skills can build an ad campaign, create a commercial, and localize ads. Those are the named examples supported by the source. The entry labels the release for all plans, but access and behavior should still be checked in the operator’s account on the test date.

Test “build an ad campaign” as a bounded package: premise, script, shot list, draft assets, and review summary. Exclude media buying, claims approval, account access, and publishing.

Agent 2.0 context

The June 25 entry describes Agent 2.0 as creating campaigns, analyzing data to improve creative, and scaling marketing across platforms, formats, and markets. The May 13 entry describes Runway Agent as a conversational creative partner for developing and producing a finished video. These dated statements give context to the later skills announcement; they do not remove the need for a task-level test.

Open the official Runway Agent workspace only after writing the brief and approval boundaries. Save the visible product label, settings, and date; do not infer a hidden model from output style.

What the changelog does not promise

The cited entry does not promise perfect factual accuracy, character continuity, rights clearance, disclosure compliance, or approval-free publishing. Absence is not proof that a feature cannot exist elsewhere; it means this changelog entry cannot support those claims.

List access limits, retention, localization, editable files, approval controls, and delivery behavior as unknown until verified.

Choose what to delegate

Delegate tasks with frozen inputs and visible acceptance rules. Retain brand, identity, legal, financial, and publishing decisions. Assign every human review in a RACI-style matrix.

Safe delegated tasks

Good candidates include transforming an approved brief into variants, organizing a shot list, adapting aspect ratios, drafting localized copy for review, and producing clearly labeled concept assets. Each output must remain reversible and unapproved until a person checks it.

If the task depends on generation rather than orchestration, test the same brief on the APOB Runway alternative page. APOB’s practical advantage is a focused creator path: the operator can move from an owned image or prompt into a controlled video test without first handing an agent responsibility for the whole campaign. That narrower surface can reduce delegation ambiguity, though only matched outputs can establish which path fits.

Human approval gates

Place gates after the brief, script, keyframe or character choice, generated draft, localization, claims review, and final export. A gate needs a named reviewer, input bundle, checklist, decision, timestamp, and rollback action.

If “polish” changes the persona, product, or claim, return to the relevant gate.

Never-delegate decisions

Keep final control of real-person identity, consent, regulated or comparative claims, licensing, disclosure, budget, account permissions, and publishing with accountable people. This test also forbids real-person impersonation and unowned brand assets.

An agent may summarize evidence; it cannot certify its own output as lawful, truthful, or approved.

Build one matched creator brief

Use the same brief across Runway AI, a focused model or tool, and any manual fallback. Lock the business goal, product facts, persona, audience, formats, and pass conditions before generation so the comparison is about workflow—not changing instructions.

Locked inputs

Copyable brief: “Create a 15-second vertical concept for an owned reusable water bottle. Audience: adult commuters. Persona: fictional adult creator, age 28, neutral wardrobe, no resemblance to a real person. Message: easy to carry and refill. Required product facts: only facts in the attached approved sheet. Tone: practical, calm, observational. No health, environmental-impact, or comparative claims. Use only attached product images and licensed music.”

Attach the approved facts, images, brand colors, pronunciation notes, prohibited claims, and aspect ratio. Version the files.

Required deliverables

Require five hooks, one chosen script, a six-shot list, one keyframe, one 15-second draft, a caption, and Spanish and Japanese draft copy for human localization review. Also require an evidence note listing source files, generated files, settings, time spent, retries, and unresolved questions.

Every path produces the same deliverables; otherwise record “not supported” or “manual step required.”

Pass/fail conditions

A pass preserves the fictional persona and product, uses only approved facts, completes the required action, fits the delivery format, and leaves no blocking visual, audio, localization, or rights issue. A failure log names the first timestamp or text span where the requirement breaks.

Allow two documented attempts per path; unlimited retries hide time and failure cost.

Audit the agent output

Review source claims and creative quality separately. A polished video can still contain an unsupported sentence; a cautious script can still show the wrong product or an inconsistent person.

Claim verification

Extract every factual sentence into a table with claim, approved source, exact support, reviewer, and verdict. If the source sheet does not support it, remove or rewrite the claim before the creative is evaluated.

An opening link is not enough; it must support the wording in context.

Character and brand continuity

Compare face, hair, body proportions, wardrobe, product geometry, logo placement, colors, and accessories across the keyframe and every shot. Review at normal speed and frame by frame around cuts or motion peaks.

Use the broader APOB AI Video Generator when the same locked brief needs a comparison across another available generation path. APOB’s advantage is rapid model comparison inside one creator workflow; it does not replace the continuity check.

Localization review

Give each locale reviewer the English source, audience, tone, prohibited claims, product names, numbers, and the intended on-screen duration. Check omissions, changed meaning, untranslated English, line breaks, and text that no longer fits.

Save corrected locale text and reviewer identity, then edit only from that approved version.

Rights and disclosure

Confirm ownership or permission for every image, video, voice, logo, typeface, and music asset. Record that the persona is fictional and not presented as a real customer. Check the platform and campaign’s current disclosure requirements with the responsible owner; this article does not provide legal advice.

Keep consent, license, disclosure, and the final asset in one evidence bundle.

Decide where Agent Skills fit

Base the conclusion on observed effort and failures in this one workflow. Record the result as of August 26, 2026 and schedule a September 2 recheck, or earlier if Runway changes the Agent or Agent Skills documentation.

Best-fit tasks

Delegate when the input is approved, output is reversible, acceptance rule is observable, and a human gate exists. Use a focused model or APOB tool when the need is primarily a controlled image or video generation step. Keep the work manual when it requires sensitive access, final authorization, or evidence the agent cannot verify.

Retest conditions

Retest when access, named skills, models, plan limits, localization behavior, editable outputs, or approval controls change. Preserve the original brief and accepted files so the new result can be compared rather than remembered.

Manual fallback

The fallback is the approved brief, source bundle, and checklist—not an instruction to “do it manually.” Assign the producer, editor, localization reviewer, rights owner, and publisher. Rehearse the handoff once before scaling.

Final decision checklist: confirm named capabilities against the current changelog; lock a fictional adult persona and owned assets; cap attempts; record every output and retry; verify claims; review continuity and every locale; confirm rights and disclosure; compare human effort; select agent, focused tool, or manual path; then obtain final sign-off. The test succeeds when responsibility is clearer, even if the decision is not to delegate.

Sources

• Runway product changelog — official launch record

• Runway Agent — official workspace

• Runway — official site

• Runway AI video generation alternative — APOB AI

• AI Video Generator — APOB AI