Four services, one pipeline. Hephaestus forges the agent from a goal spec. Nyx is the blueprint every agent inherits — Strands on AWS AgentCore, runnable locally. Proteus tunes the agent's objective against a dataset, co-creating an adversarial LLM-as-judge when the target cannot be scored with a metric. Panoptes watches everything — one view per run, one view across all experiments.
GoalSpec into a governed agent. 14 phases with evaluator gates
(pass / retry / fail / escalate). Loads the Nyx blueprint, picks a pattern,
generates Strands code, hands a genome to Proteus, then packages the winning config
for deployment. Drafts the adversarial LLM-as-judge when objectives are fuzzy.
Strands tools, risk tiers, circuit breakers,
gated actions, morning briefings. Targets AWS AgentCore in prod
and runs locally — non-negotiable, because Proteus must optimize them on dev loops.
LLM-as-judge and runs an adversarial loop
until the judge is stable and the candidate converges.