Swamp Ushers in the Second AI Frontier with Behavior-Centric Design

New architecture backed by $7 million in seed funding led by Amplify Partners replaces fragile, token-hungry agent designs with reusable behaviors shown to work reliably every time at 1/10 of the cost and up to 6X speed

SAN FRANCISCO--(BUSINESS WIRE)--Swamp Club today announces the general availability of Swamp, a platform for building AI agents with behavior-centric design, a new architectural approach that solves the reliability and cost problems that prevent enterprises from scaling AI in production. The industry's current agent-centric approach uses frontier models to decide the control flow of our applications, and the results are fragile, token-hungry, and difficult to scale. Swamp instead uses frontier intelligence to translate a developer's intent into reusable, deterministic behaviors that run the same way every time. New AI agents are built in hours, use less than 1/10 the tokens of LLMs or other frameworks, and run everywhere from a developer's laptop to the most demanding enterprise environments, up to 6X faster.

Gartner® Research predicts that over 40% of agentic AI projects will be canceled by 2027¹, and, according to McKinsey, only 37% of enterprises using AI can show positive EBIT impact². The root cause is architectural. The prevailing approach to building AI agents puts the LLM in charge of control flow, deciding what the program should do next at every step. As these systems grow more complex, they become fragile, expensive, and impossible for a team to reproduce. Swamp’s behavior-centric design provides the answer the industry needs as it shifts from experimentation to production results.

"Our marketers know exactly who they want to reach, and they shouldn't need a DevOps engineer to act on it. Swamp lets us give that control back to them, safely. They ask for what they need, AI works out what they mean, and Swamp carries it out within the guardrails we create,” said Adam Roberts, SVP Data Science, Analytics, & Insights at Xero. “Work that used to take hours or days now takes minutes. We started with one prototype for marketing, and now other teams at Xero are building on the same platform."

Behavior-centric vs. Agent-centric Design

In agent-centric design, the LLM reasons about what to do, calls a tool, adds the result to its context, and reasons again. Each added capability makes the system more fragile because correctness depends on the model making the right probabilistic decision at every step. They also get more expensive -- each turn requires more and more context, consuming expensive frontier intelligence to make every decision. To compensate, teams add validation agents and retry loops, which further multiply token consumption without eliminating the underlying fragility.

Behavior-centric design separates understanding from execution. Frontier intelligence translates a developer's intent into reusable behaviors, each a distinct architectural unit with typed inputs and outputs, versioned data schemas, and isolated credential access. Those behaviors compose into deterministic workflows that call an LLM only when a specific step requires reasoning. It uses the developer’s AI agent or coding harness to turn their intent into reliable behaviors that can be maintained by frontier intelligence.

A controlled comparison built with open-source code illustrates the difference. A trip-planning agent built with LangChain's agent-centric pattern consumed 24,611 input tokens. The same agent built with Swamp's behavior-centric approach consumed 1,183 input tokens, roughly 14x fewer, at one-sixth the cost per plan. Both repositories are published on GitHub.

What Swamp Does

Swamp works with a developer's preferred harness to turn intent into well-factored, modular behaviors. It works with any service, API, or database. It provides a scalable, queryable data layer that stays consistent across devices and team members. It manages credentials through encrypted vaults that agents can't see. And it embeds production-grade security, observability, and control. Every behavior a team creates adds to a shared corpus that other teams can extend and reuse, turning individual wins into organizational capabilities.

Swamp usage has grown quickly, driven by engineers teaching each other what works, in public. Since the first prototype shipped in March 2026, the community has published more than 1,500 extensions with more than 100 independent write-ups on what they have built. One engineer moved 11 automation flows off n8n in a single day, putting every workflow and configuration change under version control. Another built a 15-model Databricks integration covering jobs, pipelines, SQL warehouses, and Unity Catalog in about four hours, testing every model against a live Databricks workspace.

"The industry has spent years consumed by developments from the frontier labs, with engineers furiously trying to harness them for practical ends," said Adam Jacob, co-founder and CEO of Swamp Club. "The results have too frequently been a failure. When the LLM controls what happens next, the system gets more fragile and more expensive with every capability you add. Existing frameworks compound the problem because they were built around that same agent-centric assumption. Swamp is built around a different one - use frontier intelligence to understand what someone wants, turn it into reusable behaviors, and then execute it as deterministic code that works every time. The result is reliable, auditable, and cost-effective AI systems that solve the real problems engineers face."

"AI can now build almost anything. The hard part is trusting that code to be reliable and repeatable in production. Swamp lets developers ship agentic automation that works, not sometimes works," said Sunil Dhaliwal, General Partner at Amplify Partners and lead investor in Swamp Club. "Deterministic, cost-effective AI is the biggest opportunity in venture, and the Swamp Club team are the best people in the world to make it happen. It's no surprise that the Swamp movement is taking off."

"I used Swamp to launch our production SaaS platform at Bitbison by simply describing the architecture we wanted, and letting Swamp build it all for me," said Samy Al Bahra, CEO and Co-Founder of Bitbison. "What would have taken weeks of effort, we handled in a day and created a flexible system all of our engineers could work with. Swamp has saved me headcount by enabling our engineers to safely manage and extend our infrastructure within guardrails, not to mention all the time I got back."

In Short:

• The problem: The prevailing approach to AI agents puts the LLM in charge of control flow, making systems fragile, token-hungry, and impossible to reproduce across teams. Gartner predicts over 40% of agentic AI projects will be canceled by 2027.
• The architecture: Behavior-centric design decomposes work into reusable behaviors with typed inputs and outputs, versioned data, and isolated credentials, then composes them into deterministic workflows that call frontier intelligence only when a step requires it.
• The evidence: A production code review workflow saw an 8x reduction in token consumption while running faster. A controlled trip planner comparison showed 14x fewer input tokens and 6x lower cost. At Xero, engineering adopted Swamp after marketing had already built behaviors through the software approval chain, producing the cross-departmental reuse the architecture is designed to enable. Production SaaS platforms launched in a day. Software factories that enable individual developers to safely and securely ship hundreds of times a week.
• What developers are building: Software factories, infrastructure automation, complex lead scoring, biology experiment management, and more. Over 1,500 community-built extensions since March 2026.

"We call it Swamp Club because it's a thriving community of engineers working together to build the practical frontier of AI agents, not just software," said Nick Stinemates, VP of Business Strategy. "We're constantly learning from each other, building on top of one another's experiences, and pushing each other to go further and do more. Join the club. It's time to build."

For more info:

• Behavior-Centric Design Paper
• Swamp Club
• Swamp Leaderboard
• The Feed: User-generated how-tos and “what I did with Swamp”
• Swamp Club Podcast: “AI Frontiers”

^[1]Gartner Newsroom, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, June 2025. Gartner is a trademark of Gartner, Inc. and/or its affiliates. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027

^[2] McKinsey & Company, AI is changing work. Now it has to change the organization https://www.mckinsey.com/capabilities/people-and-organization/our-insights/ai-is-changing-work-now-it-has-to-change-the-organization 

About Swamp Club

Swamp Club provides a platform to build AI agents using behavior-centric design and scale them from a developer's laptop to the most complex production environments. It is the fastest way for a developer to go from an idea to a working AI system. Founded in 2026 and headquartered in San Francisco, CA, the company is backed by $7 million in seed funding from Amplify Partners, Encoded Ventures, and select angels. More information is available at https://swamp-club.com.