[KongchangAI]
Concept目标模式 / Goal-based Agent

Goal Mode

An automated execution mode for AI Agents in which the user specifies a high-level objective and the model autonomously runs iterative cycles—executing tasks, self-verifying results, and planning subsequent steps—until the objective is fulfilled, without requiring human intervention at intermediate stages.

Core Facts

Timeline (last 90 days)

Sep 30

Goal mode is an automated implementation approach for handling complex requirements in AI agents; after the user issues a goal, the model starts a loop that automatically executes, self-reviews, and plans the next round until all rounds are complete.

Unverified50%
Sep 30

Goal mode is essentially an engineering-level encapsulation of the ReAct (Reasoning + Acting) loop.

Unverified50%
Sep 30

On top of ReAct, Goal mode adds persistent storage (SQLite for saving goals and state), budget management, and mandatory self-review validation after each round.

Unverified50%
Sep 30

The difference between normal mode and Goal mode is that normal mode requires the user to send a confirmation each round, whereas Goal mode only requires the user to issue the goal and perform the final acceptance, with all intermediate steps completed automatically in a loop.

Unverified50%
Sep 30

Using a while loop to simulate Goal mode has three fundamental differences: it cannot be stopped or started at any time, it has no self-verification function, and it cannot execute round by round.

Unverified50%
Sep 30

Creating a goal in Goal mode in Codex requires explicitly selecting "Goal mode" or informing the model in the prompt; Goal mode is not created automatically.

Unverified50%
Sep 30

A precise goal prompt should contain five elements: terminal state, evidence, boundaries, recurring tasks, and stopping rules, of which terminal state and evidence are mandatory.

Unverified50%
Sep 30

After the model in Goal mode receives the prompt, it only estimates what to do in the first round; after completing and verifying that round, it then checks whether the task is finished, and the total number of rounds is unknown in advance.

Unverified50%
Sep 30

The signal that a round has ended is that the content returned by the model no longer contains any tool calls, or the model explicitly returns "this round should end."

Unverified50%
Sep 30

The Goal mode flow provides four hooks, which subscribe to four events respectively: start, end, stop, and continue.

Unverified50%

3 more timeline events

All Facts (14)

Verified

Goal mode uses a single message as both the task initiation instruction and the completion criterion, so Codex knows not only what to do but also when it is done.

70%
Unverified

Goal mode is an automated implementation approach for handling complex requirements in AI agents; after the user issues a goal, the model starts a loop that automatically executes, self-reviews, and plans the next round until all rounds are complete.

50%
Unverified

Goal mode is essentially an engineering-level encapsulation of the ReAct (Reasoning + Acting) loop.

50%
Unverified

On top of ReAct, Goal mode adds persistent storage (SQLite for saving goals and state), budget management, and mandatory self-review validation after each round.

50%
Unverified

The difference between normal mode and Goal mode is that normal mode requires the user to send a confirmation each round, whereas Goal mode only requires the user to issue the goal and perform the final acceptance, with all intermediate steps completed automatically in a loop.

50%
Unverified

Using a while loop to simulate Goal mode has three fundamental differences: it cannot be stopped or started at any time, it has no self-verification function, and it cannot execute round by round.

50%
Unverified

Creating a goal in Goal mode in Codex requires explicitly selecting "Goal mode" or informing the model in the prompt; Goal mode is not created automatically.

50%
Unverified

A precise goal prompt should contain five elements: terminal state, evidence, boundaries, recurring tasks, and stopping rules, of which terminal state and evidence are mandatory.

50%
Unverified

After the model in Goal mode receives the prompt, it only estimates what to do in the first round; after completing and verifying that round, it then checks whether the task is finished, and the total number of rounds is unknown in advance.

50%
Unverified

The signal that a round has ended is that the content returned by the model no longer contains any tool calls, or the model explicitly returns "this round should end."

50%
Unverified

The Goal mode flow provides four hooks, which subscribe to four events respectively: start, end, stop, and continue.

50%
Unverified

For subscription users in Goal mode, as long as there is remaining quota in the window the current round can start; a slight overage is deducted in the next round, and once the weekly quota is exhausted the task automatically stops and resumes when the quota is restored.

50%
Unverified

Goal mode includes five states: over-limit, completed, Blocked, paused, and deleted; three consecutive failures within a single turn cause the task to be marked as blocked.

50%
Unverified

The core difference between Goal mode and Plan mode is that Goal iterates in rounds with verification after each round, while Plan mode generates a plan and executes it all at once without verification.

50%

Source Articles