1 minute read

execute, then delegate

while most AI agent frameworks today are conversation-obsessed and production-hostile, deltagency cuts through the noise by implementing a task-based, Ollama-native AI agent framework with zero bloat.
in contrast to layers of abstractions of other frameworks, the goal is to provide a clean, simple interface to work with and make the workflow as deterministic as possible.

vision

deltagency aims to provide:

  • modular components instead of rigid hierarchies
  • optimization for input → output workflows, not endless chat
  • direct Ollama integration with full API control
  • declarative YAML agent definitions
  • minimal dependencies, maximum control

architecture

deltagency is built with a simple but structured approach.
its architecture combines clean design layers with a flexible relay pattern for orchestration.

design layers

the core of deltagency is organized in clear layers.
each layer has a single responsibility, making agents easier to build, debug, and extend.

  • communication bridge → direct Ollama API integration
  • agent executor → task execution
  • relay nodes → composition of workflow made of multiple agents
  • relay runner → the control loop that executes the nodes

the relay pattern

traditional frameworks mix execution with decision-making, creating debugging nightmares
deltagency introduces the relay pattern, which separates concerns:

  • execute → agent completes its specialized task
  • delegate → agent passes results or decides the next step
  • clean → context is released, resources freed
# build any orchestration pattern
A ↔ B (consultation loop) → C (finalize)
Router → {SQL_Expert | XSS_Expert | Auth_Expert}
Coordinator ↔ {Worker1, Worker2, Worker3}

one pattern, infinite workflows

why it’s different

most frameworks optimize for conversations or heavy orchestration.
deltagency is different:

  • task-focused → deterministic workflows, predictable outputs
  • modular → composable, reusable components
  • ollama-native → direct API integration, no wrappers
  • lightweight → only requests, PyYAML, jsonschema
  • developer-friendly → YAML configs and a clean API

status

deltagency is under active development.
early results:

  • better modularity
  • cleaner configurations
  • equal accuracy with less overhead

code will be open-sourced once production stability is reached.

in summary

deltagency is a lightweight, task-first framework for building modular AI agents.
it focuses on deterministic workflows, direct Ollama integration, and minimal dependencies.
the goal is simple: remove overhead and give developers predictable, controllable agents.

case studies



Updated: