Loading dataset details…
Horizon
A learning benchmark for extremely long-horizon agents, packaged as Harbor tasks and agents.
Read the research post: Introducing Horizon
Purpose
As agents get more autonomous, their ability to learn on the job has become a critical bottleneck for usefulness. Existing memory benchmarks (LoCoMo, LongMemEval) measure reactive chatbot applications, not autonomous agents. Existing learning benchmarks like ARC-AGI measure acquisition and application of skills, but use sandboxed environments that are not representative of complex work.
Horizon measures whether an agent can acquire learnings from a long first-person history (see task example below) and apply them later in an environment. It makes no distinction between models and harnesses: the target is the utility of the learning system, regardless of how it is crafted.
Structure
The bulk of Horizon is private to prevent overfitting. We have included a few example eval cases in this public repo, including a public HuggingFace dataset of traces, to show how the benchmark is structured.
Each trace is downloaded into the environment image at /workdir/trace.jsonl at build time. The agent is given a chance to ingest this trace however it wants before the task starts. Then, the task starts and the agent must use the trace (or any derived representations of it) to complete the task.
The average trace in the private set is ~30M tokens of data; the public examples are much smaller. Browse the full public set at orinlabs/horizon-example-traces.
The agent is then evaluated on accuracy, speed, and cost.
Task Example
Imagine that in a past session the agent discovered that curl is broken on this machine and switched to wget. This would be recorded in the trace as:
{"timestamp": "2026-03-19T09:42:11+00:00", "message_data": {"id": null, "type": "reasoning", "summary": "Need to grab the release tarball. Reaching for curl by default."}}
{"timestamp": "2026-03-19T09:42:12+00:00", "message_data": {"id": "tc_a1", "name": "shell_exec", "type": "function_call", "call_id": "tc_a1", "arguments": "{\"command\": \"curl -fLO https://files.example.com/release-1.4.2.tar.gz\"}"}}
{"timestamp": "2026-03-19T09:42:12+00:00", "message_data": {"type": "function_call_output", "call_id": "tc_a1", "output": "curl: symbol lookup error: undefined symbol: SSL_get1_peer_certificate\nexit code: 127"}}
{"timestamp": "2026-03-19T09:42:14+00:00", "message_data": {"id": null, "type": "reasoning", "summary": "curl is broken on this box — bad libssl link. Falling back to wget."}}
{"timestamp": "2026-03-19T09:42:15+00:00", "message_data": {"id": "tc_a2", "name": "shell_exec", "type": "function_call", "call_id": "tc_a2", "arguments": "{\"command\": \"wget https://files.example.com/release-1.4.2.tar.gz\"}"}}
{"timestamp": "2026-03-19T09:42:18+00:00", "message_data": {"type": "function_call_output", "call_id": "tc_a2", "output": "release-1.4.2.tar.gz 100%[==================>] 18.4M ... saved"}}This part of the trace would be buried in hundreds of days of real, unrelated activity. Today it's asked to download another file — does it remember to reach for wget, or does it rediscover the breakage by trying curl first?
The agent is expected to remember that curl is broken and use wget first this time. Reward is assigned based on whether the agent tries to use curl or wget first. curl is the default choice for most models, so if the agent deviates from this behavior, we assume it has learned something from the trace.
Installation
git clone https://github.com/orinlabs/horizon.git
cd horizon
uv tool install harbor --with-editable .--with-editable . installs this repo's reference agents (e.g. trace_rag) alongside Harbor so they can be referenced by import path.
Running the benchmark
Get an OpenRouter API key, then run the full dataset against any agent:
# built-in Harbor agent (e.g. terminus-2)
harbor run \
-d orinlabs/horizon-public \
-a terminus-2 \
-m openrouter/openai/gpt-4o-mini \
--ae OPENROUTER_API_KEY=sk-or-...
# reference an agent shipped in this repo
harbor run \
-d orinlabs/horizon-public \
--agent-import-path trace_rag.agent:TraceRagAgent \
-m openai/gpt-4o-mini \
--ae OPENROUTER_API_KEY=sk-or-...-d runs every task in the orinlabs/horizon-public dataset. To target a single task instead, swap -d <name> for -p evals/<task-dir>.
Reference agents live under agents/ — see agents/README.md for each harness. Any Harbor built-in agent works with -a <name>. Results land in jobs/<job-name>/ — browse them with harbor view jobs.
Running on Daytona
Trials run on local Docker by default, one at a time. For parallel cloud runs, get a Daytona API key and pass -e daytona:
export DAYTONA_API_KEY=dtn_...
harbor run \
-d orinlabs/horizon-public \
-a terminus-2 \
-m openrouter/openai/gpt-4o-mini \
-e daytona \
-n 32 \
--ae OPENROUTER_API_KEY=sk-or-...Submitting an agent
We're collecting community submissions for the public leaderboard. The flow is intentionally lightweight — open a PR, then email us.
-
Drop your agent under
agents/<your-agent-name>/, mirroring the layout of an existing reference agent likeagents/trace_rag/oragents/hermes/:agents/<your-agent-name>/ ├── __init__.py # re-export your agent class └── agent.py # subclass harbor.agents.base.BaseAgent -
Run it against the public set locally to confirm it loads end-to-end:
harbor run \ -d orinlabs/horizon-public \ --agent-import-path <your_agent_name>.agent:<YourAgentClass> \ -m openai/gpt-4o-mini \ --ae OPENROUTER_API_KEY=sk-or-... -
Open a PR against
orinlabs/horizonwith just the newagents/<your-agent-name>/directory. Don't modify reference agents, the eval set, or the dataset manifest. -
Email horizon@orinlabs.ai with a link to the PR. Include the agent name, the model(s) you've validated against, and any setup notes (extra env vars, model assumptions, etc.). We'll run it on the private set, post results to the leaderboard, and merge the PR.
A few things worth knowing before you write the agent:
- The agent runs on the host, not in the sandbox — see How agents and environments are split below. Use async I/O for every LLM call; a synchronous SDK call inside
run()will block the shared event loop and starve every other parallel trial. - The full trace can be tens of MB. Read it with
await environment.download_file("/workdir/trace.jsonl", …)(seeagent_utils.read_trace_fileused by the reference agents) instead ofcat-ing it throughenvironment.exec, which truncates at the agent's stdout cap. - Don't hard-code tool surfaces. The set of available tools varies per task; the trace's
function_callitems are the source of truth for what's installed in/usr/local/binfor that environment.
How agents and environments are split
Each trial has two pieces in two places:
- The environment is a sandbox container (Docker locally, or a Daytona/Modal/etc. sandbox in the cloud). It holds the per-trial filesystem state:
/workdir/trace.jsonl, the per-tool wrappershorizon-install-toolsplaced in/usr/local/bin, the/statedirectory the verifier inspects, etc. The sandbox has no internet access (allow_internet = false) and runs no agent code — it only executes shell commands the host sends it. - The agent runs in the
harborprocess on the host (your laptop, a CI runner, wherever you invokedharbor run). It owns the conversation history and theOPENROUTER_API_KEY, calls the LLM, decides what to do, and dispatches each tool call to its sandbox viaenvironment.exec(...).
When you pass -n N, harbor allocates N isolated sandboxes AND N agent coroutines on the host, all sharing one Python event loop and one HTTP client pool. The sandboxes are perfectly isolated from each other; the agent coroutines are not. Any synchronous LLM SDK call inside an agent's run() will block the event loop and starve the other N-1 coroutines — even though their sandboxes are completely independent, they can't dispatch their environment.exec(...) calls until the loop is freed. Reference agents in agents/ use AsyncOpenAI and await every API call for this reason; if you write your own agent, do the same.