Agent
What should the agent do?
Load a workspace to start: agentic-harness serve --workspace workspaces/hello
Memory
Long-term memory: facts the agent chose to keep across conversations, one list per profile. Each new task starts with its profile's facts in the agent's instructions. Forget one and it is gone from the next task on.
Online Evals
Live answers are scored after you see them: never blocking, never sent back to the agent. Code metrics are exact and free, on every answer. LLM as Judge metrics need one small paid judge call per answer, so they run only while the judge is on. This profile's conversations, newest first.
Off: only the free code metrics run. On: one judge call per answer (about $0.002 with Flash-Lite).
Metrics
Answers
Offline Evals Synthetic Data
Agent
Answers the test questions
$0
Judge
Grades every answer
$0
You Pay
$0
Running
- Agent Answering
- Judge Grading
- Done
Metrics
Test Results
Past Executions
Golden Dataset Synthetic Data
Checks
Server … Key …
Not run yet.