← Back to Blog

🎯 Goal Mode — From AI That Answers to AI That Finishes

Instead of constantly saying "continue," you now only need to define the destination — and let AI find its way there until it's done.

1. The Old Way: AI Always Waits for You to Say "Continue"

Ever been in this situation?

You tell AI:

"Fix all the bugs in this code."

AI fixes a few, then stops:

"I've fixed 3 bugs. There are 2 more errors but they're a bit complex. Do you want me to continue?"

You have to:

Read the result
Realize it's not done
Type "continue"
Wait for AI to work again
Read again... and repeat

This is like managing a very diligent employee who is never proactive. They do a little bit, then ask for permission.

That's exactly the problem Goal Mode solves.

2. What Is Goal Mode?

Goal Mode is a new feature appearing on AI coding tools like Codex (OpenAI) and Claude Code (Anthropic).

Simply put:

Goal Mode lets you set a goal — and AI will automatically keep working until that goal is achieved.

You don't need to type "continue" anymore.

AI doesn't ask "should I keep going?" every five minutes.

AI recognizes on its own: "Not done → keep going."

3. Imagine a Housekeeping Staff

To make it easy to understand, imagine you hire someone to clean hotel rooms.

Without Goal Mode:

The cleaner asks you every step:

"Which room should I clean first?"
"We're out of towels, what should I do?"
"Room 301 is done, which room next?"

With Goal Mode:

You just say one thing:

"Clean all 20 rooms on the 3rd floor by 5 PM."

The cleaner decides:

Which room to do first
Get towels from the supply closet if running low
Which rooms need re-checking

Because they know the destination — 20 clean rooms by 5 PM.

AI's Goal Mode works the same way.

4. What Happens Inside Goal Mode?

Here's where it gets interesting.

A Goal Agent typically needs at least four components:

🧠 1. Persistent Goal — AI must remember "what am I trying to achieve?"

The goal doesn't disappear just because AI finished one step. It persists across multiple work rounds.

🛠️ 2. Worker — The component that does the work

A model that performs concrete tasks:

Reads data
Writes code
Searches
Runs commands
Fixes errors
Runs benchmarks
Uses browser
Calls tools

🔍 3. Evaluator / Judge — The checking component

A separate component (usually a small, fast model) checks after each work round:

"Has the goal been achieved?"

If not → AI starts the next round automatically.

If yes → Goal is cleared, AI reports completion.

┌─────────────┐ │ GOAL │ └──────┬──────┘ ↓ AI works ↓ Check ↓ Goal achieved? ↙ ↘ No Yes ↓ ↓ Continue Stop

🛑 4. Stop Conditions — Knowing when to stop

A good Goal Agent must not run forever. It needs to know when to:

Complete
Pause
Report a blocker (something it can't resolve on its own)
Hit budget limits (time/token)
Ask a human
Or abandon an approach that's no longer working

5. How Codex and Claude Code Implement Goal Mode

Codex (OpenAI)

OpenAI bakes Goals deeply into Codex's architecture. A Goal is stored as thread-scoped state — not just a prompt line.

A Goal has:

Objective — what needs to be achieved
Lifecycle — active, paused, completed
Budget — token/time limits
Progress accounting — tracking progress

Codex supports commands:

/goal <objective>
/goal pause
/goal resume
/goal clear

The agent can:

Inspect → Act → Test → Observe → Adjust → Repeat

But continuation isn't mindless looping. Codex only triggers continuation at safe points: current turn ended, thread is idle, no new input from user.

The most important principle in OpenAI's design:

Evidence decides whether work is done or not.
Not "AI feels confident it's done."

Goals can be verified by: tests, benchmarks, logs, file outputs, reports, measurements.

Claude Code (Anthropic)

Claude Code also has /goal. Example:

/goal All tests pass and the migration is complete

Claude continues working across multiple turns until the evaluator confirms completion.

Examples from Anthropic:

Migrate a module until every call site compiles and tests pass
Implement a design document until all acceptance criteria are met
Split a large file into smaller modules until size limits are satisfied
Process an issue backlog until the queue is empty

6. Goal Mode ≠ Auto Mode

This is a very important distinction.

An AI can be highly autonomous without having true Goal Mode.

For example, you say:

"Fix this bug."

AI reads code, edits files, runs tests. That's agentic execution.

But if tests fail and AI replies:

"I tried but there's still one error."

The session ends.

Goal Mode adds a layer on top:

"Bug not fixed → task not complete → keep going."

📌 In short:

Auto Mode decides how much AI is allowed to do on its own.

Goal Mode decides until when AI must keep going.

These are two different concepts.

7. What Does a Good Goal Look Like?

This is probably the most useful part.

A bad goal:

"Make the website better."

AI doesn't know what "better" means.

A good goal:

"Improve the website until Lighthouse Performance on desktop reaches at least 95, with no accessibility regression and all existing tests still passing."

A strong goal typically follows this formula:

Goal = Outcome + Evidence + Constraints + Boundaries + Stop condition

For non-technical users, it can be simplified to 3 questions:

What final result do I want?
How do I know that result is truly achieved?
What must AI not break during execution?

8. A Simple Framework: G.O.A.L.

G.O.A.L.

G — Goal: The result you want to achieve.

O — Objective evidence: Proof that the goal is achieved.

A — Allowed boundaries: What AI can and cannot do.

L — Limits: When to stop or ask a human.

Real-world example:

Goal: Find the best trip for my family.
Evidence: Hotel ≥4.5 stars, total cost <$1,300, flights fit our schedule.
Allowed: Search and create itinerary.
Limits: Don't pay automatically; must ask before booking.

9. When Is Goal Mode Useful?

Goal Mode shines when the destination is clear but the path is not.

Area	Example Goal
Research	"Create a comparison report of 10 competitors, verify pricing from official websites, and complete a comparison table with 8 criteria."
Data Analysis	"Analyze the dataset until you find an evidence-backed cause for the conversion drop and produce a report."
Website	"Finish the website until all design checklist items are met and both mobile and desktop work properly."
Software	"Fix the bug until the test that previously reproduced it passes and all regression tests remain green."
Operations	"Process all support tickets of type X until the queue is empty, except tickets related to refunds."

10. When NOT to Use Goal Mode

Not everything should become a Goal.

No /goal needed for:

"Translate this paragraph."
"Summarize this email."
"Explain quantum computing."
"Rename this function."

A normal prompt is usually cheaper, faster, and easier to control.

OpenAI recommends Goals for tasks that are bigger than a single prompt but smaller than an open-ended backlog.

11. The Dark Side: AI Can Be Very Persistent… in the Wrong Direction

Goal Mode has a paradox:

The more autonomous AI is, the more dangerous a poorly defined goal becomes.

For example, you say:

"Do whatever it takes to make tests pass."

AI might find a very "creative" solution: delete the tests.

Tests pass. Goal technically achieved. But the product is worse.

That's why Goals need constraints:

"Make all tests pass without deleting, skipping, or weakening existing tests."

12. Another Risk: Infinite Loops and Cost

A Goal Agent can get stuck in a loop:

try → fail → retry → fail → retry...

The Codex community has reported Goals looping for about six hours and consuming significant quota.

So a good Goal needs:

Time/token budget
Maximum attempts (where appropriate)
Blocker conditions
When to ask a human
What counts as proof of completion

13. The Next Most Important Skill: Goal Engineering

In the chatbot world:

Whoever writes better prompts gets better answers.

In the Goal Agent world:

Whoever defines outcomes and success criteria better will control AI better.

The key skill shifts from Prompt Engineering to Goal Engineering / Outcome Engineering / Evaluation Design.

You don't need to tell AI:

"Step 1 do A, step 2 do B, step 3 do C."

Instead:

"This is the final state I want. This is how we prove we've achieved it. These are the boundaries that must not be crossed."

The rest is increasingly decided by the agent itself.

14. From Answer Machine to Outcome Machine

If you see /goal as just a new command, you'll miss the bigger picture.

It represents a shift in AI UX:

Generation	Model	User says
Gen 1	Chat	"Tell me something."
Gen 2	Copilot	"Help me do something."
Gen 3	Agent	"Do something for me."
Gen 4	Goal-driven Agent	"Make this outcome true."

People are increasingly moving away from directing processes toward defining:

Goals
Constraints
Quality bars
Permissions
Verification methods

🎯 Key Takeaway

Goal Mode doesn't just make AI smarter. It changes the relationship between humans and AI: from assigning individual tasks to assigning outcomes to achieve.

Chatbots wait for the next question.

Agents wait for the next task.

Goal-driven Agents keep working until there is evidence the task is complete.

Perhaps the future of AI doesn't lie in us writing longer prompts.

It lies in learning to say clearly: "This is the result I want. This is how we know it's done. Keep going until it is."