Autonomy Loop
Long-Horizon Execution
The planning and self-supervision loop that keeps an agent on a goal across hours of work: decomposition, progress tracking, and recovery when a step stalls or wedges. The hard part is finishing, not starting.
01 / THESIS
THE MODEL
CAN REASON.
THE AGENT
MUST
FINISH.
Emotions Research is an agent lab. We build the most reliable long-horizon autonomous agents: systems you can hand a sprawling, multi-day task and trust to finish it unsupervised. We do not train models. We build the agent layer on top of them, the planning, self-supervision, recovery, and verification that turn a capable model into an agent that actually completes the work.
02 / MANIFESTO
A model that follows instructions is a tool. It stops where the prompt ends. The hard problem is not generating an answer. It is holding a long task together for hours without a human catching every drift.
Frontier models can already reason, plan, and write code. Pointed at a long horizon, the agents built on them drift, wedge, hallucinate progress, and quietly fail. That is an agent problem, not a model problem, and it is the only problem we work on.
Every research track at Emotions Research moves toward agents that hold a goal across hours of unsupervised work, recover from their own mistakes, and ship a result you can review. We build the cold structure that makes long-horizon autonomy dependable.
03 / WHAT WE BUILD
Verification
Ship On Green, Not On Hope
Every change runs in an isolated git worktree and is gated on types and tests before it lands. A task is done only when it passes. A throwing or empty run is caught and re-driven, never quietly marked done.
Reliability
Fleet Observability and Self-Heal
The instruments that keep a long run alive: heartbeated brain loops, a supervisor that restarts a wedged agent, and structured telemetry across the whole fleet. You cannot trust what you cannot see.
04 / FLAGSHIP
AUTONOMOUS CODING FLEET
NYX
Nyx is our first deployed system: a local-first autonomous coding agent that plans, writes, tests, and ships software on its own machine across long, unsupervised sessions. It dispatches a fleet of parallel workers, gates output on types and tests, and delivers reviewable branches. Nyx is where our research on long-horizon autonomy stops being theory.
$ nyx "add a /health endpoint with a test"
Will live at nyx.emotionsresearch.com