AI evaluations

See how Ask Antares and Antares's standalone AI Skills perform. Open any target to inspect its runs, cases and evidence.

8
AI products
1
with a baseline
78
cases

Ask Antares

One agent, evaluated at two levels: the complete experience and each internal skill it can load.

Overall agent Ask Antares The Work-site agent that routes questions to a skill, reads the right context, calls tools and returns grounded answers or resources. 78 casesRunner connected 100%Public baseline11/11 passed
Overall agentEnd-to-end behavior across routing, retrieval, tools, answers, resources, multi-turn context and latency. 6 internal skillsFocused contracts for the skills Ask Antares loads while completing a request.

Standalone AI Skills

Products published outside Ask Antares. Each keeps its own benchmark, runner and score.

Standalone AI Skill Eynap Everything You Need as an AI PM 0 casesRunner not connected 37/37Static checks--No baseline yetNot run Standalone AI Skill Agent Memory Hub Shared memory for ChatGPT, Claude, Codex, and Claude Code 0 casesRunner not connected --No baseline yetNot run Standalone AI Skill /devloop An unattended dev loop that works your issue queue 0 casesRunner not connected --No baseline yetNot run Standalone AI Skill /iterate A human-reviewed improvement loop for other skills 0 casesRunner not connected --No baseline yetNot run Standalone AI Skill /ai-pm-resume A résumé reviewer that reads like an AI PM hiring loop 0 casesRunner not connected --No baseline yetNot run Standalone AI Skill /interview-sim Synthetic interviews that read like real transcripts 0 casesRunner not connected --No baseline yetNot run Standalone AI Skill Claude Credit Watchdog Resumes your Claude Code work the moment credits renew 0 casesRunner not connected --No baseline yetNot run

Functional dimensions

Dimensions help people design useful cases. They are not evaluation targets.

01Routing

Does the agent or product choose the right capability for the request?

02Grounding

Does the answer use the correct source and avoid unsupported claims?

03Tool use

Are the right tools called, with no unnecessary or repeated calls?

04Resource delivery

Are files and links correct, usable and free of duplicates?

05Answer quality

Is the response direct, complete, correctly scoped and well written?

06Multi-turn context

Does a follow-up keep the right context without leaking old output?

07Latency

Does the first useful output and the completed answer arrive quickly enough?

08Reliability

Does the behavior stay correct across languages, wording and repeated runs?