Unreal Agent
A new Go-based AI coding harness called Unreal Agent aims to make tool-using agents more efficient by running tools asynchronously, reducing polling overhead, and cutting token usage compared to popular setups like Codex and Pi. Commenters are broadly positive on the async, actor-like orchestration model and compare it with other harnesses and techniques (e.g., programmatic tool calling, decision models, and fractal tool discovery), while questioning how much real-world speed and cost benefit it delivers. Many also criticize the branding, noting likely trademark conflict with Epic’s Unreal Engine and the confusion it creates for developers expecting game-related tooling.
Async tools & token/cost savings
- Several commenters struggle to see how async tool calls directly save tokens, since models aren’t “running” while tools execute.
- Explanations focus on:
- Reducing wasteful polling loops in some harnesses (e.g., Codex repeatedly checking task status).
- Avoiding cache expiry while long-running tools execute, which can force expensive context reloads.
- Doing “more tool work per model turn,” reducing cache reads and total turns.
- Some note similar token savings achieved by tuning timeouts and wait behavior in existing harness configs.
Harness architecture & async/actor models
- Strong enthusiasm for treating agents as asynchronous processes/actors rather than linear chat sessions.
- Proposed model: message-passing between “actors” (filesystem, tools, other agents), with notifications instead of polling and support for cross-device routing.
- Several see this as a natural fit with Erlang/Elixir-style supervision and failure handling, essentially an “agent OS”.
Comparison to existing harnesses & techniques
- Multiple frameworks already support async tools or related ideas: OpenAI’s async tool calling, Claude Code, Crush, DSH’s “programmatic tool calling”, and several open-source harnesses.
- Some say Unreal’s approach is close to what others have been doing since early 2024, though newer models are better trained for async patterns.
- Claims that techniques like Unreal’s can be combined with cost-optimizing harnesses (e.g., maki.sh).
Skepticism and open questions
- Doubts about real-world benefit vs. “appearance of speed,” since agents often must wait on tool output to proceed.
- Prior experiences where agents call async tools but then immediately “wait” with long timeouts, causing slowdowns or hangs.
- Concern that forbidding sub-agents or sticking to CLI-centric SDKs may hurt long-horizon or complex tasks.
Naming and trademark concerns
- Many readers initially assume this is an Unreal Engine-related agent.
- Widespread view that the “Unreal” branding is confusing and likely to attract a trademark challenge from the game engine’s owner.
- Some worry this legal risk could impact the sustainability of the project for serious users.
Usability, docs, and ecosystem issues
- Complaints that the README lacks a clear “getting started” path, though others prefer minimal, high-signal docs.
- Questions about support for subscriptions, OAuth, GUIs/TUIs, and compatibility with existing subs (e.g., Codex).
- Broader frustration that experimentation with new harness ideas is costly under API pricing, limiting innovation despite clear low-hanging fruit.