HAI AgentGraph

Agent state graph is an integral of the trace

Todo

Infer

1. checkSchemaType → merged | langsmith | langfuse

     IF merged:
       → mergedToGraph() (existing, done)

     IF langsmith/langfuse:
       ↓
  2. saveTraceToFile(/tmp/trace.json)
       ↓
  3. jqQuery tool - explore & extract per step:
     - inferStateType
     - inferSubject
     - extractData
     - extractTools
     - inferPrompts
     - extractTimestamp
       ↓
  4. Map phase:
     - Direct inference OR
     - Generate function → double-check → use

  Ready to start implementing? I'll add minimal tools:
  1. createSaveTraceToFileTool()
  2. createJqQueryTool()
  3. Update checkSchemaType to detect langsmith/langfuse
  4. Update prompt for new flow

• inferStateType

• inferSubject

• extractData

• extractTools

• inferPrompts

• extractTimestamp

Integration - leftshift

data schmea

visualization

how to test more accruately

Perturbation Testing

agent mode

Problems

• AgentSteer mlflow trace and unilever mlflow trace are different
• That is fine but what is the real original from from mlflow

• Source code limitation

AgentSteer Structure

1. MLflow Tract

2. Basic Graph

3. Detailed Graph

4. Reactflow Graph

5. Visualization

Instruction

write a @traverse-agent.ts script which can be run 


pnpm shell .src/agentgraph/traverse-agent.ts ˜˜.json

using vercel ai sdk (refer @exposure-assessment.ts for how to using)
of mastra agent

it reads provided json and acces

data.orchestrator_traces.messages (only using that)

and than traverse the message while update updating the output graph and memory

@merged_orchestrator_6909ed23e4486dfa127548a1.json (32-44)  

1. for example reading the first message
it update global memory stack - question
make an first edge and node
the edge's subject attribute is user and target node is Asked status for example

@merged_orchestrator_6909ed23e4486dfa127548a1.json (45-73) 
2. the new edge and status node planned is created. the edge's subject might be competitor 360 agent and explanation might be added to memory or not.
obviously plan should be added on global memory stack

@merged_orchestrator_6909ed23e4486dfa127548a1.json (73-101) 
3.  new edge whith user subject and new state node approved

@merged_orchestrator_6909ed23e4486dfa127548a1.json (103-115) 
4. ares subject is moving to the edge and status to research state

@merged_orchestrator_6909ed23e4486dfa127548a1.json (116-128) 
5. self referencings edge with ares subject 

@merged_orchestrator_6909ed23e4486dfa127548a1.json (129-141) 
6. new state and node retrievd state with ared subject edge
@merged_orchestrator_6909ed23e4486dfa127548a1.json (141-160) 
7. new reported state node with subject node reporter. if the memory size is huge don't need to store the full context and even no need to update memory **if it does not help future graph construction or helful to prevention of duplicated node or edge**


rules
- the travese agent is for construct state graph for agent trace, traversing and generating a graph
- single message can create single state node and edge
- the traverse agent have a context of what he is doing, global memory yand constructed graph
- the user or agent is a executor of the state change (edge); a edge's attribute has user or specific agent or subject 

- the agent are given with empty graph or pre-constructed graph. if the pre-constructed graph the traverse agent basially traverse them with changin state based on the edges. However if some trace components are not interpreted by pre-given graph or current graph (the agent only knows the current graph; no concept of pre existing). update or create new edge or node to traverse correctly for example
- all graph is dag
case 1

if use rejected the plan the agent might create state rejected from planned with user subject or create an edge to asked state based on implementation

case 2 

if planning is asked to another agent, off to compertitor agent , it should create another edge but not state and just with another subject 


did you got it or is theeere another question  

first you should sort by message.created_at

answer
1. graph schema
{}  and state node type should be dynamically selected not fixed.  memory (context) only contains - current piece of trace, current graph, global memory and instruction rules etc

memory is jsut key value stroage please do not make it as a fixed type please
for edge remove explanation and timestamp
type is stirng  remove timestamp and message id


you don't get it now. We are constructing  **generatlized graph which are capable to explain every traces generated from a same agent system** we are not generating a graph for a single trace. 

edge has another attribution with optional as 

system_prompt
instruction_prompt
potentially with template string if you are confident with semantic variable name in this agent system's context
Also you can update the useful global memory
- system description
with suspectful tone since the agent is only seeing a peice of the trace . we means it need to be udpated along the trace

the  output grpah schema is

{
states
edges? or transitions? what is the term using is finite state machine context
memory{record string stirng?}
}


there may be a several tool to generate graph and managing a graph


- create_state
- create_edge?


when ai agent call function, the improtant thing is ai agent should not rewreuite the whole graph. the agent just wirte a state json or edge json then it pass to the tool and update the runtime object managed. (returns success or not only not the full graph inserted to context since agent already have it)

create_memory
update memroy
delete memryo is also same
only key and value is generated by llms as toksn and the return is success or not

any more question?

this is not for orchestrator. this is for any agent system. change the schema name
use fs promises not using import * path either
never identify message.type manually inside code. the agent jsut read the full json of the message and no unnecessary fine tune to this trace schema


the code only knows the message is a list of unknown

you can define schama of this tracce as merged azure trace but never fine tune on that

Merged architecture

• conversation {}

• turns []

• messages []

• plans []

• events []

• files = []

• prompt_cards []

• user_feedbacks = []

• status_store = []

• stats {}

• natural_language_query

• competitor_df

• chat_history = []

• standalone_query

• rephrased_query

Orchestrator architecture

• events

• turns

• plans

• prompt_cards

• feedbacks = []

• execution_trackers

• files = []

1. Technical Excellence

2. Poster & Presentation

3. Real-world Impact

• Track A: Performance Optimization / Robustness

• Track B: Observability / Explainability

• Track C: Attack Discovery / Methodology