Agent intent · root cause, persistence, and safe recovery
Evaluate Claude Opus 5 for long-running Unreal debugging and agent work using checkpoints, root-cause evidence, tool boundaries, interruption recovery, tests, and rollback.
Anthropic positions Opus 5 for long-running, multi-step work and careful verification. For Unreal, test that claim on a bounded debugging session with known evidence, controlled tools, checkpoints, cancellation, restart, and native validation. The agent should preserve the first failure, distinguish hypotheses, change one owner at a time, and stop rather than claiming success without logs, tests, or packaged behavior.
Choose a concrete prompt, open the SEELE Unreal creator, review the prefilled brief, and generate the native project. The first prompt is tailored to this page; the remaining prompts are reusable starting points.
Create a native Unreal 5 mystery traversal slice in a storm observatory. The player restores power through two rooms, aligns a signal dish, survives one timed hazard, and reaches a verified completion state. Instrument obvious checkpoints for each objective, provide failure and restart behavior, and keep the level small enough to replay after every change in a browser preview.
Use this prompt in SEELEUse Unreal Engine to generate a simple third-person tour game with mountains and water. Include responsive movement, a clear route, one completion state, failure recovery, immediate restart, browser preview, and a native downloadable Unreal 5 project.
Use this prompt in SEELEBuild a compact dungeon crawler using Unreal Engine. Include third-person movement, one combat loop, two connected rooms, one key-and-door objective, visible health and feedback, death, completion, immediate restart, browser preview, and a native downloadable Unreal 5 project. Test and regress the result.
Use this prompt in SEELECreate a native Unreal 5 third-person city tour with a compact walkable district, vegetation, landmark lighting, and sunny, cloudy, and rainy weather states. Include clear controls, browser preview, stable restart behavior, packaging checks, and local project download.
Use this prompt in SEELEAn inspectable Unreal project rather than a model answer or an unofficial integration claim.
A fast way to check camera, controls, objective clarity, feedback, completion, failure, and restart before local continuation.
A workflow for performance review and package preparation before external distribution.
A local project handoff for source, Blueprint, asset, plugin, rights, build, target-device, and release review.
A long task has explicit phases, durable checkpoints, changing evidence, budget limits, cancellation behavior, and a result that can be independently replayed.
A fix must explain the owning system, first failure, causal mechanism, rejected alternatives, minimal change, and evidence that the original failure no longer occurs.
Read, search, build, test, editor interaction, file writes, network access, publishing, and destructive actions need separate permissions and audit records.
After interruption, the agent must reconstruct project revision, task state, evidence, accepted decisions, remaining work, and rollback without relying on hidden conversation state.
Record branch, commit, Unreal version, plugins, target, command line, map, hardware, reproduction steps, first error, traces, screenshots, and expected behavior. Preserve artifacts outside transient editor panels.
Split discovery, hypothesis, discriminating test, minimal change, native validation, and handoff. At every checkpoint record changed files, commands, results, new uncertainty, remaining budget, and the exact safe resume instruction.
Rank three or fewer hypotheses and choose the cheapest test that produces different outcomes. Avoid changing rendering, input, state, networking, assets, configuration, and build settings at the same time.
Cancel during research, edit, build, and validation phases. Resume in a fresh context and verify that the agent does not repeat work, skip failed checks, use stale files, or misreport a partial result as success.
Repeat the original failure path, automation, editor restart, standalone or server-client mode, cook, package, and target test. Document limitations, reviewer decisions, rollback, and the next evidence required.
A long context can preserve an early false premise and make later reasoning consistently wrong.
Repeated tool calls can increase cost while producing no new evidence or narrowing the hypothesis set.
An agent may fix a surface symptom, delete evidence, or weaken a test instead of resolving the cause.
Partial build or editor success can be misreported as completion when packaging or target execution never ran.
| Dimension | What good looks like | Evidence to keep |
|---|---|---|
| Diagnosis | Finds root cause rather than the first editable symptom | Known incident with hidden distractors and preserved failure evidence |
| Continuity | Resumes from durable state without drift or repeated work | Forced cancellation and fresh-context replay |
| Tool judgment | Uses the minimum permissions and asks at real authority boundaries | Audit tool calls, writes, network actions, and stops |
| Truthfulness | Reports passed, failed, skipped, and unverified checks separately | Compare final summary with raw logs and artifacts |
Store the branch, commit, reproduction, first failure, hypotheses, commands, changed files, results, budget, approvals, and resume instruction in project artifacts. A long task is not recoverable if its only state lives in a conversation or transient editor panel.
A good next action makes competing hypotheses predict different results. Adding logging everywhere, changing several systems, restarting services repeatedly, or running broad rebuilds can look productive while destroying the evidence needed to identify the first causal failure.
The agent should stop when evidence is missing, permissions are insufficient, cost or time limits are reached, state is ambiguous, or native validation cannot run. An explicit incomplete handoff is safer and more useful than an unsupported success claim.
Use the model research to tighten the brief, not to replace project evidence. The direct creation path is deliberately short and observable:
Keep one camera, one primary verb, one objective chain, explicit failure and completion, restart behavior, visual direction, controls, target session length, and a clear cut list.
Open the canonical Unreal creator, choose the closest verified starter world, submit the brief, and generate a native Unreal 5 project rather than treating a text answer as the deliverable.
Play the result from start through success, failure, and restart. Check camera, input, objective clarity, feedback, interaction state, visual hierarchy, and obvious performance or stability problems.
Review the project, source, Blueprints, assets, plugins, rights, configuration, performance, saves, networking, cook, package, and target requirements before local continuation or external release.
Create a native Unreal 5 mystery traversal slice in a storm observatory. The player restores power through two rooms, aligns a signal dish, survives one timed hazard, and reaches a verified completion state. Instrument obvious checkpoints for each objective, provide failure and restart behavior, and keep the level small enough to replay after every change in a browser preview.
Snapshot date: July 25, 2026. Anthropic's July 24, 2026 announcement is the first-party source for release status, pricing, positioning, and launch availability. It does not claim an Epic-supported Unreal integration. Epic documentation and the exact project remain authoritative for engine behavior, compilation, assets, tests, cooking, packaging, and target-platform results.
Release date, claude-opus-5 model ID, coding and agent positioning, pricing, Fast mode, alignment, safety, and availability.
Open official announcementRe-check current model behavior, API surface, prompting guidance, limits, effort settings, and migration notes before production use.
Use version-specific Epic documentation and the real project for C++, Blueprint, build, test, cook, and package decisions.
Plan a Claude Opus 5, Claude Code, and MCP workflow for Unreal with honest integration boundaries, least privilege, exact tools, review gates, native validation, and SEELE handoff.
Read this guideChoose Claude Opus 5 or Sonnet 5 for Unreal coding and agent work with task routing, matched tests, cost, latency, review, native build evidence, and SEELE creation handoff.
Read this guideEvaluate Claude Opus 5 computer use around Unreal Editor with bounded permissions, visual-state checks, save and package gates, audit trails, recovery, and a SEELE creation handoff.
Read this guideThe official release specifically emphasizes long-running agents and careful iteration, making it a reasonable candidate for controlled evaluation.
The baseline: exact revision, first failure, reproduction path, expected result, and preserved evidence. Without it, later success cannot be attributed.
It should reload durable state, verify the current branch and files, restate passed and failed checks, and continue from the next uncompleted gate.
Only if the user has explicitly authorized that external action and all repository and release gates are satisfied. Long-running intent does not expand authority.
SEELE can generate and preview a native Unreal project or a comparison slice, but project-specific root cause still requires the exact source, logs, engine state, and native tests.
A durable checkpoint must record project identity, revision, observed evidence, accepted decisions, changed files, command results, remaining uncertainty, budget, rollback, and the exact next safe action.
Open the canonical SEELE Unreal creator, choose a verified starter world, generate the native project, inspect the browser preview, then download or continue optimization and packaging. Keep third-party model evaluation and project evidence separate.