← all publishers

AgentX-ai

@agentx-ai source repo

Report this profile Claim this profile

• 4 skills
• 0 followers
• 11 hours ago last updated

1. ▌
   Eval Fix · agentx-ai bundle

   Turn an AgentX self-host evaluation into a triaged code fix and a re-run against the same dataset, so a before-and-after comparison means something. Use whenever someone has an evaluation on a local AgentX self-host engine (AgentX-trace-eval, normally http://localhost:4700) and wants to know what to actually change in the code, or wants to re-run an evaluation to compare scores. Also use when a request mentions a self-host evaluation id or dataset id, an agent that scored badly, an AI Analysis or judge findings from the Evaluate tab, or the question "the report tells me what is wrong with the answers but not what to fix in my code". The core move is triaging code-blind judge recommendations against the real source instead of applying them literally.
   0 installs
2. ▌
   Run Eval · agentx-ai bundle

   Run an AgentX self-host evaluation of the local agent: pick or create the dataset and grading config (templates, an existing id, a CSV/file of Q&A, or cases curated from the agent's own live traces), write the harness that calls the agent per case with a linked trace, execute the run, and hand back the score, the browser report, and the /eval-fix command. Use whenever someone wants to evaluate their agent on a self-host engine (AgentX-trace-eval, normally http://localhost:4700), asks to "run an eval", "score my agent", "create a dataset", "test my agent against a dataset", or has traced runs and wants to know how good they are. The deliverable is a committed eval harness plus a finished, analyzed run - the exact thing /eval-fix picks up.
   0 installs
3. ▌
   Instrument · agentx-ai bundle

   Set an existing Python agent up on AgentX, end to end: read the project API key off the engine, write .env.agentx, install agentx-python from PyPI, initialise the SDK once, instrument the calls that are worth tracing, and then prove it by running the agent and reading its own traces back. Use whenever someone wants to set up, initialise, add tracing to, or monitor their own agent with AgentX; asks to "trace my agent", "instrument this repo", "hook this up to AgentX", "set up AgentX here", or "get my runs into Live Traces"; or has a self-host engine (AgentX-trace-eval, normally http://localhost:4700) and nothing reporting into it yet. Also use when traces are configured but not arriving, since the SDK fails silently by design. The core move is one span where the run begins plus one line of framework auto-instrumentation - not a decorator on every function - and it is not finished until a trace has been fetched back.
   0 installs
4. ▌
   Auto Improve · agentx-ai bundle

   Turn an AgentX improvement report - production failures a human CONFIRMED during signal review, clustered into issues with recommendations - into triaged fixes applied to the agent's source code. Use whenever someone has an improvement report id from a self-host engine (AgentX-trace-eval, normally http://localhost:4700), asks to "apply the improvement report", "fix what the reviewers confirmed", or mentions the dashboard's Insights > Auto-improve surface. This is the ONLINE counterpart of /eval-fix: the evidence is live traffic a human vouched for, not an offline dataset run, and the core move is the same - triage code-blind recommendations against the real source instead of applying them literally.
   0 installs