Recoleta
GitHub
Source note

New AI Agent Architecture to fix LLM deviations and token costs

Authors
nexcatara
Source
Hacker News
Published
Collection
Software Intelligence
AI AgentsWorkflow AutomationLLM Tool UseState MachinesMCP
Open GitHubMarkdown source

Summary

BotCircuits proposes an agent design where a deterministic workflow engine controls multi-step flow and the LLM handles reasoning and tool calls inside each step. The excerpt is a product README, so it gives architecture and usage details but no benchmark evidence.

Problem

  • LLM-driven agents can drift during long tasks because the model decides both what to do and how to route the process.
  • Multi-step automation can waste tokens when the model re-evaluates control flow at every step.
  • Teams need agents that can run repeatable workflows through CLI, HTTP, chat, and scheduled jobs while keeping tool execution visible and configurable.

Approach

  • Workflows are JSON records under .botcircuits/workflows/; after a build step, each workflow becomes a callable tool.
  • The LLM performs each agentAction, while the engine follows start, next, and compiled branch conditions.
  • Natural-language branch conditions are compiled into typed choices[] entries with operators and values, and flow.variables tells the runtime how to coerce inputs before branching.
  • Skills are disk folders with SKILL.md instructions and optional allowed-tools, so repeatable behavior can be added without changing the system prompt.
  • MCP servers, built-in file/shell/code tools, a FastAPI gateway, and messaging channels extend where the agent can act.

Results

  • No benchmark results are provided for token cost, task success, latency, or deviation rate.
  • The main concrete claim is architectural: runtime branching can happen without another LLM call after conditions are built.
  • The README shows support for 3 LLM providers: Anthropic, OpenAI, and Gemini.
  • Workflow authoring supports at least 2 step types: start and agentAction; branching uses conditions plus next fields.
  • The gateway can route at least 4 channel types in one FastAPI process: WhatsApp, Slack, generic webhook, and cron.
  • The example workflow has 11 steps: step_1 through step_10 plus end, with early termination controlled by end_id.

Link