Skip to main content

Example: Helpdesk Agent

This example builds a customer support agent that handles ticket construction tasks. It uses memory to get better over time: the first ticket takes 14 planned steps, but by the 20th similar request, the agent follows a proven pattern in 3 seconds.

Setup​

from masar import MasarClient

client = MasarClient()

Step 1: Recall Past Experience​

A new request arrives: "Set up a ticket routing system with SLA tracking."

memories = client.memory.recall(
    context="ticket routing system with SLA tracking",
    domain="helpdesk",
    limit=5
)

if memories.pattern:
    print(f"Found pattern: {memories.pattern.name}")
    print(f"Success rate: {memories.pattern.success_rate:.0%}")
    print(f"Based on {memories.total_matches} similar episodes")

On the first run, no memories exist. On later runs, the agent might get:

Found pattern: ticket-routing-sla
Success rate: 94%
Based on 12 similar episodes

Step 2: Plan the Build​

plan = client.plan_instructions(
    current={"name": "HelpdeskApp", "orbitals": []},
    goal="std-helpdesk",
    domain="helpdesk"
)

print(f"{plan.total_instructions} instructions across {plan.levels} levels")
for instr in plan.instructions[:5]:
    print(f"  L{instr.level}: {instr.action}({instr.params})")

Step 3: Execute With Your LLM​

schema = {"name": "HelpdeskApp", "orbitals": []}

for instr in plan.instructions:
    prompt = f"Apply this to the schema: {instr.action}({instr.params})"
    schema = your_llm.execute(prompt, schema=schema)

    # Verify after each level completes
    if instr == plan.last_instruction_at_level(instr.level):
        check = client.verify(schema=schema)
        if not check.valid:
            repairs = client.rank_edits(schema=schema, errors=check.predicted_errors)
            schema = your_llm.apply_repairs(schema, repairs.suggestions[0])

Step 4: Final Verification​

final_check = client.verify(schema=schema)
errors = client.error_check(schema=schema)

print(f"Valid: {final_check.valid} ({final_check.probability:.0%})")
print(f"Predicted errors: {errors.top_errors}")

Step 5: Store the Episode​

client.memory.store(
    domain="helpdesk",
    context="ticket routing system with SLA tracking",
    schema=schema,
    actions_taken=[i.action for i in plan.instructions],
    outcome="success" if final_check.valid else "failure",
    metadata={"goal": "std-helpdesk", "steps": plan.total_instructions}
)

What Happens Over Time​

After 20 helpdesk tasks, the agent's recall consistently returns a pattern. Planning confirms the pattern is still valid for the specific request, and execution follows the known-good path. Tasks that took 14 LLM calls on day one take 3-4 calls by week two, because the agent skips exploratory steps and goes straight to what works.

Next Steps​