Better Agent

BetterAgent is a hosted agent layer that attaches to existing Next.js codebases without new infrastructure. The workflow is three commands: `betteragent discover` walks your project and lists every server action, API route, and exported handler; you select which ones the agent may call and write descriptions the model uses to decide when to invoke them; `betteragent sync` diffs against the vendor’s server and ships the approved tool manifest. At runtime the agent reads data through GET routes, mutates state through server actions, and can steer the UI through client actions — three primitives, one protocol. Routes are locked at deploy time; the model sees only the names and schemas you approved.

The differentiating feature is auth forwarding. The vendor’s docs describe the SDK dispatching server-action calls so that your existing session context, revalidation logic, and bearer tokens pass through unchanged. This means the agent operates as the authenticated end-user, not as a separate service principal — no new IAM roles, no token-exchange middleware, no shadow auth path to audit.

BetterAgent fits Next.js SaaS teams that have already shipped a working product and want to add an agent surface without rebuilding the backend. It breaks when the agent needs to branch conditionally across tool calls based on runtime state — approved handlers are static, so multi-step conditional logic has to live inside your own handler code, not in the agent’s decision layer. Teams whose agents need to span multiple services or run tasks outside the Next.js server boundary will hit the edges of what the platform was designed for and are the most likely to evaluate a general-purpose agent framework instead.

UI components follow the shadcn registry pattern: install with `npx betteragent add sidebar` (or popup, inline-bar, cmd-k), and the code lands in your repo, inheriting your existing shadcn theme tokens. Observability is built into the hosted dashboard — every run records route, tool calls, token count, latency, and spend, queryable by run ID — so debugging a failed agent call looks like reading an HTTP access log rather than tracing a model.