AI agents run your marketing campaigns now. The stack.

Twelve months ago, "AI agent runs your marketing campaign" was a slide in a venture deck. The slides ended with a roadmap, not a product. In May 2026 the slides ended differently. The product shipped. Three of them shipped simultaneously, a

Twelve months ago, "AI agent runs your marketing campaign" was a slide in a venture deck. The slides ended with a roadmap, not a product.

In May 2026 the slides ended differently. The product shipped. Three of them shipped simultaneously, actually — Hermes Agent from Nous Research, Seedance 2.0 from ByteDance via Higgsfield's MCP server, and the maturity of Claude Code's orchestration loop — and the combination quietly made the fully autonomous marketing campaign economically rational for the first time.

We've now run six end-to-end campaigns through the agentic stack across the businesses we operate. The setup is roughly the same each time. The operator brief is a paragraph. The agent decomposes it, generates the creative, runs the distribution, measures the response, iterates inside 72 hours. Operator attention per campaign: 2–4 hours, mostly spent reviewing the hero asset. AI agent marketing automation in 2026 isn't a productivity tool layered on top of the marketing team's workflow — it's a replacement for the parts of the marketing workflow that used to need three people, with the parts that still need a human now visibly separated from the parts that don't.

This is the stack, the layer-by-layer breakdown, the operator math, and the honest list of what still needs a human in the seat.

The orchestration layer that wasn't possible 12 months ago

Three things had to be true for the autonomous marketing campaign to ship as a real product, not a demo. As of Q2 2026, all three are true at production grade.

A video model good enough for paid distribution. Seedance 2.0 cleared that bar in February 2026. Multimodal text-plus-image-plus-audio input, up to 2K resolution, native synchronized audio in a single generation pass, ELO 1,351 for image-to-video — first globally, ahead of Kling 3.0, Google Veo 3, and Runway Gen-4.5. The 2025 video models could generate clips; Seedance 2.0 generates ad-quality clips with brand consistency from reference images. That's the bar.

An agent that can hold a campaign brief in memory across sessions. Hermes Agent's /goal feature, released the same month, was the part that fixed the orchestration layer. Prior agent frameworks could execute a task. Hermes can hold a multi-week campaign brief, remember the prior week's results, and decompose next week's plan against the cumulative state. Marketing campaigns are inherently multi-session work — the agent that forgets between sessions isn't useful for the job.

An operator surface that lets a human intervene cleanly. Claude Code matured into the actual driver's seat for this kind of work in late 2025, with Claude Code as the spot you sit when you need to read a decision, override a tactic, or rebuild a slot in the campaign without restarting the agent.

Take any one of the three away and the stack doesn't work. With all three, the stack works for the first time. The 12-month window matters here — operators who tried this in early 2025 burned weeks on systems that couldn't ship. Operators starting today walk into a stack that's already integrated.

The agentic marketing campaign, layer by layer

A campaign in our pipeline runs through five layers. Each layer has an owner — sometimes an agent, sometimes an external tool, occasionally a human. The orchestration is built so each layer can fail without taking the rest of the campaign down.

Layer 1: brief and strategy. Hermes Agent takes the operator's one-paragraph brief and decomposes it into a campaign plan — audience, hook, asset count, distribution channels, success metric, budget envelope.

Layer 2: creative generation. Seedance 2.0 (and Nano Banana Pro for stills, where applicable) generates the visual assets. The agent picks the shots, writes the prompts, calls the model, evaluates the output, regenerates the misses.

Layer 3: assembly and orchestration. Claude Code stitches the campaign together — final cuts, copy generation per platform, asset variants for vertical/horizontal/square, naming and tagging for tracking.

Layer 4: distribution. Direct API calls into Meta Marketing API, TikTok Spark Ads, YouTube Shorts API, and the relevant CMS endpoints. The agent handles the upload, the targeting setup, and the budget pacing.

Layer 5: measurement and iteration. OpenClaw factory-style background agents pull performance data on a schedule, surface anomalies, propose next-batch creative variants based on what's winning.

The pattern is intentional. Each layer is replaceable. If Seedance 2 has capacity issues on a given day, the pipeline falls back to Kling 3.0 or Veo 3.1 without breaking. If Hermes hits a hallucination on the strategy layer, Claude Code catches it on the orchestration review pass. The reliability comes from the redundancy, not from any single component being perfect.

Layer 1: Hermes /goal does the campaign decomposition

The hardest part of the autonomous campaign isn't generating the creative. It's reading a one-paragraph brief and turning it into a coherent multi-asset, multi-channel, multi-week plan. This is what Hermes Agent's /goal feature was built for.

The brief into Hermes looks like this:

Launch the new winter accessory bundle next Tuesday. Target parents in DACH, 30–45, two-kid households. Hero video for Meta, three shorts for TikTok, one editorial photo set for Instagram, one launch email. Budget €600 paid spend across the first 72 hours. Goal: ROAS above 2.4.

Hermes returns a structured plan — five assets with shot lists, copy outlines, distribution timeline, budget pacing rules, and a measurement schedule. The decomposition is the part that used to be a marketing manager's full day of work; Hermes does it in roughly 90 seconds with the persistent memory of past campaigns informing the decisions.

The /goal part matters more than the model behind it. Other agent frameworks can execute single tasks well. Hermes is built around the multi-week-goal abstraction natively. The persistent memory ("we ran a similar campaign in October — the parent-voice creative outperformed the kid-voice creative 1.8x") is what makes the second campaign cheaper than the first one.

Layer 2: Seedance 2 + Higgsfield run the visuals

Once the brief is decomposed, the creative generation layer takes over. For video, this is where Seedance 2.0 does the work.

We access it through Higgsfield's MCP server, which sits in front of Seedance 2.0 plus 15+ other video and image models. Hermes calls Higgsfield's API directly, picks the model based on the shot requirement (Seedance for product-consistent shots, Veo 3.1 for cinematic shots, Kling 3.0 for motion-heavy shots, Nano Banana Pro for stills), generates the asset, evaluates output quality, regenerates if necessary.

For a 5-asset campaign, this layer typically:

• Runs 8–14 Seedance generations to land 5 keepers (first-pass keep rate ~50–65%)
• Runs 3–6 Nano Banana Pro generations for thumbnails and stills
• Costs €15–35 in model credits all-in
• Completes in 25–45 minutes of unattended runtime

The Seedance 2.0 capability that makes this work for paid marketing — not just organic content — is reference-driven generation. The agent passes a product reference image plus a brand-tone reference clip, and Seedance produces output that's actually consistent with the brand identity. The 2025 generation of video models couldn't do this. The output was visually impressive but brand-inconsistent, which made it unusable for paid distribution.

This is the same kind of leverage we documented in the 50-variants-per-week ad creative pipeline, but at full-campaign scope rather than ad-variant scope.

Layer 3: Claude Code is the operator's surface

Claude Code is where a human actually sits during the campaign. The agent runs in the background; the operator opens Claude Code, asks "what's the state of the winter-bundle campaign?", and Claude Code summarizes the asset bundle, the distribution status, the in-flight metrics, and any pending decisions.

The interaction model is conversational. "Show me the three hero variants Seedance returned" — Claude Code surfaces them. "Reject variant 2, regenerate with more warmth in the lighting" — Claude Code calls Seedance with the modified prompt. "Push the launch back two days — the cultural moment isn't right" — Claude Code reschedules across all five distribution endpoints and adjusts Hermes' goal timeline.

This layer is the thing that operators who try to build agent-only pipelines miss. Pure-agent pipelines fail because there's no clean place for a human to intervene when judgement is needed. Claude Code as the operator surface gives you both — the campaign runs autonomously by default; the human steps in cleanly when a call has to be made.

The architectural pattern is the same one we've documented running 33 autonomous agents in the OpenClaw factory — narrow agents doing narrow work, with the operator's surface designed for the moments when the agents need a referee.

Layer 4: OpenClaw handles the recurring back-office work

A marketing campaign isn't a one-shot delivery. It generates ongoing work — daily performance reports, account-health monitoring, scheduling tweaks, ad-creative refreshes when fatigue sets in, customer-service ticket routing when the launch creative attracts inquiries.

This is the work OpenClaw factory's long-running agents are built for. The pattern we documented previously — narrow charters, persistent memory, redundant orchestration — applies cleanly here. For an active marketing campaign, we typically run 3–5 OpenClaw agents in the background:

• A performance-tracker agent that pulls metrics from Meta, TikTok, GA4 every six hours and pushes anomalies to Claude Code
• A creative-fatigue agent that watches the CPM curve and proposes refresh creative when the lift starts decaying
• An inquiry-router agent that catches incoming campaign-driven customer messages and routes by intent (sales / support / refund)
• A budget-pacing agent that adjusts spend distribution across channels based on conversion velocity

None of these need a human in the loop on a daily basis. They run for the duration of the campaign and report up to Claude Code when something requires the operator's attention. This is the back-office layer of the autonomous marketing campaign — the part nobody pitches because nobody finds it sexy, but the part that determines whether the campaign actually compounds beyond the first 48 hours.

Layer 5: the autonomous measurement and iteration loop

The closing layer is where the campaign starts to learn.

Every 24 hours, OpenClaw's performance-tracker dumps the day's metrics into a structured report. Hermes Agent reads the report, compares against the original campaign goal, and proposes adjustments — new creative variants targeting the segment that overperformed, budget reallocation away from the channel that's underperforming, copy iterations on the assets approaching fatigue.

The proposals get surfaced through Claude Code for the operator to approve. Most of them get approved within a 5-minute review. The next batch of Seedance 2 generations is in the pipeline within an hour of the operator's approval. The campaign iterates on a 36–72 hour cadence — faster than any human-led campaign team we've ever benchmarked, including agencies running daily standups.

This is the loop that makes the autonomous campaign genuinely better than a human-only campaign, not just cheaper. The iteration cycle compresses from weekly (small agency) to sub-daily (agentic pipeline). The lift compounds; campaigns that go through three full iteration cycles outperform campaigns that ship once and run by 35–60% on ROAS in our data.

The operator math: a 5-asset campaign costs €40–80 and runs in 4 hours

Concrete numbers from the six campaigns we've shipped through the stack.

Line item	Cost per campaign
Seedance 2.0 video credits (8–14 generations)	€15–35
Nano Banana Pro stills (3–6 generations)	€2–5
Hermes Agent token costs	€5–15
Claude Code orchestration tokens	€10–20
Background OpenClaw agent compute	€3–8
Total tooling cost	€35–83
Operator attention (campaign lifecycle)	2–4 hours
Paid distribution budget	separate, channel-dependent

The €35–83 line excludes paid distribution budget — that's the channel spend, separate from the production stack cost. The comparable production cost for the same 5-asset campaign quoted by small agencies in 2024 was €1,500–4,000. The cost compression is 20–80x on the production side.

Operator attention of 2–4 hours per campaign is the more important number. The cost of agency production was the visible one; the cost of operator attention to brief, review, approve, and iterate was the invisible one. Agencies typically required 8–15 hours of operator attention per campaign (briefing meetings, creative review rounds, approvals, post-mortem). The agentic stack compresses both lines.

Three out of the six campaigns we ran through this stack outperformed their human-produced predecessors on ROAS. One was flat. Two underperformed — and in both cases the issue was upstream of the agents (brand-voice drift in one, an off-cycle launch timing in the other), not the stack itself. The win rate at full statistical noise is 50–60% in our data — meaningfully positive, not transformatively so. The real win is the compression of cost and time, with quality at parity.

What still needs a human — and what doesn't anymore

A short list of what humans still own, and a parallel list of what they don't.

Humans still own:

The opening 8 seconds of the hero video. The taste call. Agents converge to a generic-good opening; the moment that lands viscerally requires a human eye.

The decision to break the format because the brand voice demands it. The agent optimizes within the brief; departing from the brief is a judgement call.

The final approval before paid distribution starts. The cost of an off-tone campaign in 2026 is the same as in 2024 — the brand damage doesn't care that an agent shipped it.

The reading of the cultural moment. An automated campaign cannot know that a national tragedy makes the planned launch tone-deaf. The operator's job, increasingly, is sitting in the cultural moment and pulling the lever to delay or rework.

Humans no longer own:

Cut selection within a video. Done by the agent against the brief.

Copy iteration on existing creative. Done by Hermes faster and cheaper than a copywriter would.

Performance tracking and anomaly detection. Done by OpenClaw's background agents on a 6-hour cadence.

Budget pacing and channel allocation inside an approved budget. Done by the agentic budget-pacer.

Asset variant generation for vertical/horizontal/square. Done at zero marginal cost by the orchestration layer.

The pattern is the same one we've seen across revenue automation more broadly: mechanical work goes to agents, judgement work stays human, and the value of the judgement work goes up, not down, because there's now more of it visible per operator hour.

The agents aren't better marketers than us. They're better at the parts of marketing that look like assembly. The parts that look like judgement are still ours — and our judgement is now visibly worth what we used to charge for the assembly plus the judgement together.

— our paid lead, after the third campaign through the agentic stack outperformed the human-produced baseline

How to ship version one this weekend

If you're an operator wanting to build the autonomous marketing campaign stack, the entry point is smaller than the system diagram looks.

Friday afternoon: install the access layer. Higgsfield MCP for video and image generation (5 minutes if you have a Higgsfield account; 20 minutes if you need to sign up first). Claude Code as your operator surface (already installed if you're reading this). Skip Hermes for the first version — Claude Code can hold the brief in a single session, which is enough to ship version one.

Friday evening: write your campaign brief as a single Markdown file. Audience, hook, asset count, channels, budget, success metric. One paragraph each. Save in your project as briefs/2026-05-XX-launch.md.

Saturday morning: have Claude Code decompose the brief and generate the first asset bundle. Five Seedance 2 generations, three Nano Banana Pro stills, copy variants per channel. Total time: 60–90 minutes of operator attention plus 25–45 minutes of unattended generation runtime.

Saturday afternoon: ship to one channel. Pick the cheapest distribution channel (usually TikTok organic, or a small Meta ad budget under €50). Run for 24 hours. Don't worry about iteration yet — just prove the pipeline lands.

Sunday: review what shipped, write the second brief. Now decompose with the prior week's data as context. This is when the workflow starts to compound.

The full stack — Hermes for goal decomposition, OpenClaw for back-office, multi-channel distribution — is the second iteration, not the first. Operators who try to ship the full stack on day one consistently fail because they spend the weekend wiring agents instead of running a campaign. Ship version one in 48 hours, iterate from there. Build the pipeline against an actual campaign, not in isolation.