Swamp Club emerges with $7M seed: the 'behavior-centric' way to build AI agents that use 14x fewer tokens
YouTube is bringing its AI-disclosure habit to music: the company announced Wednesday that YouTube Music will tag tracks created “primarily using generative AI” right on the player page, with a tap-to-expand explanation of what the model actually did. The labels will also expand to the main YouTube app — and YouTube says its own detection tools can apply them even when nobody disclosed anything.
Swamp Club, a San Francisco startup founded in 2026, announced Wednesday the general availability of its Swamp platform alongside a $7 million seed round led by Amplify Partners, with Encoded Ventures and select angel investors participating. The pitch: a new way to build AI agents it calls “behavior-centric design” — and a controlled comparison showing a trip-planning agent built on Swamp used roughly 14 times fewer input tokens than the same agent built the conventional way.
Behaviors, not loops
In the prevailing “agent-centric” pattern, the language model sits in the driver’s seat, deciding control flow at every step — reasoning over prompts, tool calls, and intermediate results again and again. Swamp flips that: frontier models translate developer intent into reusable “behaviors” — typed inputs and outputs, versioned data schemas, isolated credential access — that compose into deterministic workflows. The LLM gets called only when a specific step actually requires reasoning.
That’s the bet Swamp’s co-founder and CEO Adam Jacob is making about where the agentic AI market is heading. Sunil Dhaliwal, General Partner at Amplify Partners, put it in venture terms in the release, calling deterministic, cost-effective AI “the biggest opportunity in venture.”
The 14x number
The release’s headline claim is a head-to-head: a trip-planning agent built with LangChain’s agent-centric pattern consumed 24,611 input tokens, while the same agent built on Swamp used 1,183 — roughly 14x fewer, the company says, at about one-sixth the cost per plan. Both repositories are published on GitHub for anyone to inspect.
Swamp’s broader claims follow the same logic: agents built in hours rather than days, under one-tenth the tokens of rival frameworks, and up to 6x faster. Numbers to treat as vendor claims — but the GitHub comparison repos give skeptics something concrete to check, which is more than most agent-framework launches offer.
Who’s buying in
Since its first prototype shipped in March 2026, Swamp says it has picked up more than 1,500 community extensions and over 100 independent write-ups of projects built on the platform — a real community signal for a company that, until today, was in stealth.
On the enterprise side, early customer Xero started with a single marketing prototype and has since had other teams build on the platform. Adam Roberts, SVP of Data Science, Analytics & Insights at Xero, is quoted saying work that took hours or days now takes minutes. Bitbison CEO Samy Al Bahra says his team launched a production SaaS platform in one day on Swamp, against weeks of estimated effort.
The stakes
Swamp is framing its launch against two bleak industry datapoints it cites in the release: Gartner’s prediction that more than 40% of agentic AI projects will be canceled by 2027, and McKinsey’s finding that only 37% of enterprises using AI can show a positive EBIT impact. If agent-centric architectures really are the reason projects stall — too expensive, too flaky — a cheaper deterministic pattern has a clear audience.
Why it matters
The agentic AI boom is running into its cost-and-reliability wall in public: every enterprise pilot that balloons a token bill or hallucinates a control-flow decision feeds the Gartner 40% prediction. Swamp’s answer is architectural, not incremental — fewer LLM calls, not cheaper ones. The 14x figure is a single benchmark and deserves independent verification, which the published comparison repos make possible. What’s not debatable: Amplify Partners and Encoded Ventures just put $7 million behind the proposition that the LLM doesn’t need to drive every step. If they’re right, the next wave of agent frameworks won’t be more autonomous — they’ll be more deterministic.