About Recurse by Synnada

Agents that build other agents are the idea behind Recurse by Synnada, a serverless harness its makers showed on 25 September 2026. The developer describes the job to a coding agent and says how success will be judged. Using Recurse, that agent writes candidate specialist agents, runs them and compares the results, repeating until one reaches the stated standard.

From a stated bar to a deployed agent

Each variant consists of a system prompt and Python functions exposed as tools, described by a manifest that fixes its identity, runtime and input and output schemas. The engine refuses a manifest it does not recognize, the input schema validates each run and applies defaults, and each successful run produces values conforming to the complete output schema. The winner can be published as a tool, as an MCP server or as a bot. Only the pertinent functions are registered as tools alongside Recurse's built-ins, which keeps the agent's toolset small.

Examples on the product page

Two examples show the range. One is a level designer that produces game levels in the developer's own format and plays each through a simulation; if pacing or reachability checks fail, the next layout changes, and only passing levels get to the editor. Another is an RNA sequence designer that proposes candidates, computes how each folds at minimum energy and corrects just the bases behind each mismatch. When a search stalls, the makers say, the parent agent swaps in a different specialist instead of offering an answer of its own.

Built from client work

Synnada says it built Recurse for its own client projects, where an agent takes a fixed kind of request and returns a fixed kind of response, above all in domains where correctness can be tested. Case studies on game level design appear on its blog from August 2026.

Billing and diligence

There is no subscription: runs are paid per request, model calls are passed through at the providers' token rates, and serverless execution is charged by the minute. New accounts start with free credit and need no card. Buyers should ask which model providers receive their data and where runs execute. The name is shared with unrelated organizations, including the Recurse Center, which is why the product is filed as Recurse by Synnada.

TrustList summary

Recurse by Synnada is a serverless harness for building and running specialist AI agents. A developer states the task and how success will be judged, and a coding agent uses Recurse to generate, run and compare variants of an agent until one meets that bar.

What Recurse by Synnada does

Agent variants: builds candidate agents from a system prompt and Python functions used as tools, then keeps the one that passes.
Manifest contracts: pins each agent in a manifest with input and output schemas, and refuses manifests it does not recognise.
Deployment: publishes the chosen agent as a tool, an MCP server or a bot.
Schema-checked output: every successful run returns values that match the full output schema.
Example agents: the site shows agents for game level design and RNA sequence design.

Who it serves

Teams with agents that need a fixed request and response, especially where a result can be checked automatically. The makers built it for their own client work, and their blog includes case studies from August 2026 on game level design.