bosquet.agent.graph — io.github.zmedelis/bosquet 2025.10.15

defagent clj macro

(defagent name graph entry-node node-map)

One-stop macro to define and run an agent To run the agent call (<name> graph entry-node nodemap)

defgraph clj macro

(defgraph name & edge-defs)

Create a loom digraph that can be run within the agent (edges + optional labels)

defnode clj macro

(defnode name args & body)

Define each node behavior get the current state and execut the body on the state The node function must always return a map with the values to be used in the future states Anything saved in the :completion key will be pushed into history. will be traversed to under the :next key The state in general is of the form {:key1 <output for node behavior> :key2 ... :history [[:node-name {:bosquet/completion....}]..] :trace [[:node1 :node2]...[:node-n :end]] ;the actual path taken by this agent run } Return any keys that you would like to return in the state that you want future nodes to use {:my-key 'some value returned by the llm or the result of a search' :completion 'the result of an llm call that will be added to history' }

resume-graph clj

(resume-graph interrupted-result updated-state)

Resume graph execution after human intervention Takes the interrupted state and updated user input to continue execution

run-graph clj

(run-graph graph-def
           node-map
           entry-node
           {:keys [trace __pos history] :as input})

Graph executor run a digraph given

• graph-def - a loom digraph, branch-conditions and functions to select branches {:graph <loom graph> :condition-maps <optional {:source-node {condition destination-node}..}> :condition-fns <optional {:source-node (fn[state] return values in condi}tion in the condtions map to branch on)}> }
• a node-map mapping the nodes to behaviors
• the entry node recording the history(completions) as it traverses the graph
  If a node returns {:interrupt true}, execution pauses and returns interrupt state