Model-owned task completion

AWorld's default agent completes a task when the model emits a normal final response without tool calls. Task correctness belongs to an independent evaluator. Response wording, length, repetition, artifact checks, and inferred goal progress do not override the model's completion decision.

The bundled agent uses success-first execution by default. Ordinary execution is observed rather than steered by the convergence protocol, but a candidate that reaches the terminal boundary enters one bounded, model-owned checkpoint, review, and (when justified) repair path. This review does not manufacture an external pass: it either returns the best current candidate, performs a bounded repair supported by observed evidence, or fails open with accurate limitations when review is unavailable. Set AWORLD_SUCCESS_FIRST_TERMINAL_GATE=false only as an explicit rollback of this terminal safeguard.

The default CLI completion mode is off. Callers that deliberately need the generic contract API can select observe or enforce and supply explicit artifact paths or validation commands. Paths inferred from task wording are published only as advisory deliverable hints; the legacy AWORLD_INFER_REQUIRED_ARTIFACTS flag does not create a contract. Workbench is a separately selected standalone Skill/CLI; it is not a core Tool or completion requirement.

Agent and built-in swarm step limits are disabled by default. Explicit step limits and generation watchdogs remain available to callers. Transport errors, interrupted streams, invalid tool arguments, cancellation, and externally supplied process deadlines still retain their existing failure semantics.

The generic ContextCompilerRuntimeConfig defaults to budget_pressure, with a checkpoint threshold at 95% of the available input budget. The bundled CLI Agent deliberately overrides that library default with adaptive so repeated operations, low-information observations, and durable progress can trigger an earlier checkpoint during long work. Both policies preserve the task, system instructions, work state, recent evidence, and complete assistant/tool groups; neither treats a checkpoint as proof of completion.

An external runner should save the final answer and artifacts, record any execution failure separately, and invoke its verifier after execution ends.

Harness boundary

Treat AWorld like any other packaged agent harness. The AWorld wheel owns task semantics: planning, tool-use strategy, collaborator selection, checkpointing, review, repair, and the decision to return a final answer. The outer harness may provide model transport, credentials, a sandbox and capability set, the task deadline, process lifecycle controls, and artifact collection. It must not silently select or disable AWorld semantic policy for an ordinary run.

When a caller supplies AWORLD_CONTROL_ROOT, framework-global configuration, skill state, installed skills, plugins, memory, and the global AWORLD.md layer resolve under that root. This isolates a packaged run from ambient home-directory state without suppressing explicit project instructions or artifacts in the task working directory. An agent.md discovered at runtime may not replace a collaborator already registered by the wheel's team.

Hidden evaluation remains outside the agent loop. A post-run verifier may score the saved answer and artifacts, but its reward, rubric-private state, and hidden checks are not an AWorld completion contract. Explicit single-feature ablations are experiments and should be labeled as harness overrides rather than treated as production defaults.

Default automation agent

The CLI loads Aworld as its sole default built-in agent. Its prompt guides planning, execution, observation, recovery, and completion of general automation tasks. The model chooses the workflow and useful checks. Durable scheduling is available only when the configured tool surface supports it.

Collaborators live in aworld-cli/src/aworld_cli/builtin_agents/smllc/optional_agents, outside the default agent discovery directory. None is created by default: the root model owns planning, execution, review, repair, and completion. This does not disable the long-running execution protocol's review/repair phase. It only means the root performs that phase without optional delegated advice. core explicitly exposes developer, evaluator, and the fresh-context, read-only verifier. Each selected collaborator inherits the active model, generation budget, and sandbox. Evaluator and verifier receive only a read-only filesystem tool allowlist; developer retains the writable implementation surface. None receives hidden reward or grader data. Their child contexts exclude the solver transcript, profiles, and mutable key-value scratch state; they cannot extend the parent deadline, and only each answer plus content-free usage/request evidence is merged back.

Unset or AWORLD_BUILTIN_SUBAGENTS=none disables every collaborator. all and the legacy auto value enable every available collaborator, while a comma-separated list selects exact names. The available names are developer, evaluator, verifier, diffusion, avatar, audio, and image. Skill activation remains independently controlled by default_enabled, --skill, and persisted user settings. The bundled long-running-agent skill is enabled by default but tells short tasks to bypass milestones and checkpoints; Workbench is bundled but disabled by default and remains available through explicit skill selection.