When should an AI agent replan?

An AI agent should replan when new information makes the current plan invalid, inefficient, blocked, or unsafe.

Replanning is useful when the next planned step is no longer the best way to reach the goal.

When a required step fails

Suppose the plan is:

1. retrieve customer data
2. generate report
3. email report

If the customer-data tool is unavailable, repeating the same plan may not help.

The agent may need to:

• use an approved fallback source
• ask the user for missing data
• delay the dependent step
• stop safely

When an assumption turns out to be wrong

Plans are often based on assumptions.

For example:

Assumption: The requested product is in stock.

If the inventory check returns zero, the remaining plan should change.

The agent may now:

• search another warehouse
• find alternatives
• ask the user to choose

When the user changes a constraint

A user may change:

• budget
• deadline
• destination
• preferred option
• required output

If the new constraint affects future steps, the agent should update the remaining plan.

It should not continue following instructions that are no longer relevant.

When required information is missing

Suppose the next step needs a customer ID, but none is available.

The agent may need to insert a new step:

find customer → obtain customer_id → continue

Missing dependencies are a strong signal that the plan needs adjustment.

When permissions or approvals change

An action may turn out to require:

• additional permission
• human approval
• a different role
• tenant-specific authorization

The plan should be updated to include the required approval or choose a permitted alternative.

When a better valid path becomes available

Replanning is not only for failures.

Suppose the agent originally planned three API calls, but a newly available tool can safely retrieve the required information in one call.

If the alternative is permitted and reliable, the agent may simplify the remaining plan.

When the environment changes

External systems can change while an agent is working.

Examples include:

• a flight becomes unavailable
• a meeting slot is taken
• an API goes offline
• a file is updated
• inventory changes
• a job finishes earlier than expected

Longer-running tasks especially benefit from checking whether the plan still matches reality.

When the agent is making no progress

Repeatedly performing the same steps without moving toward the goal is a sign that replanning may be needed.

For example:

search → same result → search → same result → search

Instead of looping, the agent should consider:

• a different tool
• a different query
• asking the user
• stopping

When an intermediate result changes the next decision

Some plans intentionally contain decision points.

For example:

1. check account status
2. decide next action

If status = active:

• continue

If status = suspended:

• follow the suspension workflow

The observation determines which part of the plan should remain.

When the current path becomes unsafe

The agent should replan if continuing would violate:

• permissions
• policy
• budget
• safety rules
• approval requirements

For example, if an operation now requires human approval, the agent should insert that approval step instead of continuing automatically.

When should the agent not replan?

Not every failure needs a new plan.

If a temporary network timeout occurs and the same action is still correct, a bounded retry may be enough.

Replanning is more appropriate when the path itself needs to change.

A useful distinction is:

Retry: Same action, another attempt.

Replan: Different remaining path.

Replanning should preserve valid completed work

If earlier steps succeeded, the agent should keep those results unless they are no longer valid.

For example:

• customer identified — completed
• report generated — completed
• email delivery failed

The agent should normally replan around delivery rather than recreate the report.

Replanning should have limits

An agent should not replan forever.

Useful boundaries include:

• maximum replan count
• time budget
• cost budget
• retry limits
• escalation rules
• stopping conditions

If the agent cannot find a valid path after several attempts, it may need to ask the user or stop safely.

A practical decision rule

Before continuing with the next step, the agent can ask:

1. Is the current goal still the same?
2. Is the next step still valid?
3. Are its dependencies satisfied?
4. Are the required tools and permissions available?
5. Has new information changed the best path?
6. Is the plan still making progress?

If the answer to one of these questions indicates the plan is no longer suitable, replanning may be appropriate.

In short

An AI agent should replan when new observations, failures, changing constraints, missing dependencies, permission changes, or stalled progress make the current path unsuitable.

It should avoid replanning for every minor issue, preserve completed work whenever possible, and use bounded replanning so the agent can adapt without falling into endless planning loops.