Profiles / unreallabsai

Unreal Agent

Unreal's runner resumes saved sessions around an asynchronous Go tool loop.

Edited by Michael Ruescher / reviewed 2026-10-13

Operator Stance / as of 2026-10-13

Use it for
Contained trials of async tool work, and Go integrations whose owners can enforce execution limits, consume operation outcomes and migrate sessions. Compare successful tasks, dollars and human interventions before adopting it.
Avoid it for
Unrestricted workstation use, unknown repositories opened with the runner, or consequential automation without a host that rejects failed responses before dispatch and recovery. Exit 0 and accepted cancellation are insufficient success checks.
Watch next
Use the open questions below as acceptance checks before expanding the trial.

Where it stands, 2026-10-13

Install v0.3.1 for a contained trial. The terminal client is interactive; the runner supports automation and resuming saved sessions. The stable 6 October 2026 release adds Homebrew packaging. Install the runner and unreal-agent terminal client with brew install unreallabsai/tap/unreal-agent. The terminal client starts fresh on every launch. The client runs shell commands in its launch directory without approvals or containment, with your environment inherited by child processes. Use a VM or container with restricted mounts, credentials and network access.

Choose a package you can deploy. On macOS, the tap clears quarantine from a package described as unsigned. If your fleet requires signing or notarization, put the release archives through your own process. Docker carries only the runner, as unrea1labs/unreal-agent:v0.3.1, with a digit one. For binaries or images, skip v0.3.0: its release run was canceled.

Measure the whole job before switching harnesses. The draw is async tool work and append-only, forkable sessions, plus grace timers that batch fast tool results. The vendor’s 22 September 2026 comparisons report lower aggregate input-token use against historical Codex 0.154.0/0.155.1 and Pi 0.85.1, with one Pi comparator mixing 0.85.1 and 0.86.1, using Unreal builds unmapped to public tags for Terminal-Bench, SWE-Atlas and DeepSWE. Savings on the public release remain unresolved.

Before adopting it, compare this release with your installed alternative on the same model, tasks and attempt count. Record successful tasks, input/output tokens, dollars and human interventions. A smaller model bill earns its keep only if setup, review and recovery leave you ahead.

Use the runner only on a workspace you trust. It loads the project’s .env before choosing endpoint and credentials. An unset endpoint override can make the OpenAI client send your existing key to the file’s chosen destination. The file can also replace SANDBOX_EGRESS_PROXY. Inspect it before supplying credentials, and enforce egress outside the process.

Reject failed responses before consequential automation. In this release, the coordinator can dispatch valid tool calls from a failed model response, including during recovery, and the runner can still exit 0.

An embedding host must reject those responses before either path dispatches. Inspect the runner’s JSONL log, which stores one JSON record per line, before counting a CI run as successful. The check comes too late to prevent a command already executed.

Choose credentials deliberately. The terminal client’s openai-codex default offers your existing Codex login and saves an accepted choice. Unreal Agent constructs its own tool loop. Tokens are read once without refresh. Renew with Codex and restart when rejected, or use unreal-agent -setup to change the saved provider choice.

Need to resume? Use the runner. v0.2.0 moved runner sessions from the workspace to a user-level state directory; an ID missing from that directory silently starts empty. Resume 0.1.x work with -session-directory pointing to <old-cwd>/.harness/sessions, or migrate the files before cleanup. For parallel Bash, test your provider-model pair: repeated native tool outputs are reported to cause recurring 400s on resume for some pairs. If that happens, start fresh or limit fan-out.

Enforce runtime, disk and total-spend limits outside the loop. v0.3.1 fixes SIGTERM cancellation, but Bash has no internal timeout or on-disk capture cap. Add command timeouts and a state-directory quota. The runner’s JSON max_output_tokens limits each request, not total spend; leave it unset for openai-codex, which rejects it.

Configure the thresholds for automatic conversation summaries when using custom models. The harness attempts automatic summaries from the last usage checkpoint and eligible history. Uncataloged models get threshold zero, so reported usage can trigger summaries once enough history exists.

Configure settings.json in the XDG config directory, normally ~/.config/unreal-agent. Test task accuracy before committing to a long run.

Disable fallback for fixed-model Claude evaluations. The Anthropic adapter requests the server-side-fallback-2026-07-01 beta by default. Set ANTHROPIC_FALLBACK=false to disable that request.

Port and test your Go host before upgrading. Restart deduplication now covers every persisted input kind; ResumeState.ExternalInputIDs becomes InputIDs. Adapt translators to tool.Result and ToLLMResult and preserve raw provider payloads for replay. Add and Cancel returning nil means acceptance; operation failures arrive on Updates. Wait for terminal outcomes and reconcile an unknown remote-cancellation result before advancing dependent work.

What is unresolved

  • Public-release savings and summary quality: the SWE-Atlas job identifies build 5cc55ebc88d3 and records no inference-cost total. A readable v0.3.1 comparison with its public build SHA, compaction settings, task outcomes before and after summaries, full costs and explained aggregation would settle what the install earns.
  • Removing the host’s failed-response check: a tagged fix and regression results covering both dispatch and recovery would establish when it can be retired. A nonzero exit-status fix alone would not settle execution authority.
  • Anthropic fallback: this review leaves the beta’s availability, served model and billing unresolved. Versioned provider documentation and a recorded request identifying the served model and its charges would settle them.
  • ChatGPT credentials: this review leaves permission for this third-party client unresolved. Applicable provider terms and explicit coverage of Unreal Agent would settle the subscription question.
  • Remote execution: the maintainer points to a remote-job abstraction. Remote authentication and cancellation recovery remain unestablished. A published manager with those contracts and demonstrated recovery outcomes would establish what its host can rely on.

Verification

open source commits / evidence floor: official docs / updated 2026-10-13

Verified claims

Featured in

Source policy: what Frontier watches and accepts as evidence

Edited and maintained by Bitter Frontier.

View source on GitHub