Errors and limits

Reference for what stops a task run, what a failure means, and what a task config is refused for.

The two budgets

Setting	Default	Range	What it bounds
`max_steps`	25	1 to 50	Model → tools → model rounds per attempt.
`timeout_seconds`	120	10 to 600	Wall-clock time for the whole run.

max_steps counts rounds, not tool calls. Four tools called in one reply cost one step. It is the runaway-loop guard — the number that stops a model looping over a search tool forever — not the time budget. timeout_seconds is the real bound. A run that ends without a result is re-prompted up to twice and each re-prompt starts the step budget over, so a task with max_steps: 25 can legitimately spend 75 rounds before anything gives up. Only the clock stops that. See running and runs. Lowering max_steps to tidy up is the most common way to break a research task: it fails with step_limit, which reads like the prompt is wrong when it is not.

`error.type`

A failed run carries error: {type, message}. type is a closed vocabulary of seven, and the point of it is that each one sends you somewhere different.

`error.type`	What it means	Whose	Fix
`missing_vars`	A `required` variable had no value and no default. Caught before anything was compiled, so nothing ran and nothing was spent.	Yours, in the request	Supply the value, or give the variable a `default`.
`no_output`	The model finished without calling `submit_result`, and the re-prompts did not change that.	Yours, in the prompt	Fix the prompt. Say what the task is for and that finishing means calling the tool. Do not raise `max_steps` — that was not the problem.
`step_limit`	Every round of an attempt was used and no result came back.	Yours, in the config	Raise `max_steps`, or narrow what the task does. Reported apart from `no_output` precisely so nobody rewrites a prompt that was never at fault.
`timeout`	`timeout_seconds` elapsed and the run was abandoned.	Yours, in the config	Raise the timeout, or cut the work. A slow tool is worth timing on its own first.
`provider_error`	The model provider or an MCP server failed.	Neither	Not yours to fix in the config. Retry it. A `fallback` on the LLM covers the model half.
`configuration`	Something the task needs is missing at run time: a provider key you have not added, a tool that was deleted or unpublished since the task was saved, an integration whose credential no longer resolves.	Yours, in the workspace	The message names it. Add the key under provider keys, republish the tool, or reconnect the integration.
`platform`	The run failed for a reason that is not about your task.	Ours	Retry, and tell us at hello@talqing.com if it persists.

A failed run is still a 200 with status: "failed" — the run happened, and its trace, tokens and cost are on the response. The 4xx cases below are the ones where no run happens at all.

Refused when you run

POST /v1/tasks/{task_id}/runs checks the input contract before it starts anything, and returns a 400 rather than recording a failed run.

Status	Cause	Message
`400`	A value for a name the task does not declare	`this task does not declare a variable named 'emial'`
`400`	A `required` variable with no value and no default	`'email' is required and has no default`
`422`	A `vars` value that is not a string	Substitution is textual; send `"007"`, not `7`.
`403`	The viewer role	Running a task is a write.
`404`	No such task in this region	Tasks are regional — check the base URL.

The 400 bodies use the standard envelope, with every problem listed at once:

{
  "detail": {
    "message": "these values do not match the task's variables",
    "errors": ["this task does not declare a variable named 'emial'"]
  }
}

Refused when you save

There is no publish step, so every check runs at save, including the ones an agent draft is allowed to defer. An invalid config is a 400 listing every problem, and nothing is stored — which is what makes “a task that exists can always run” true.

Refusal	Message shape
No provider key for the model. The one an agent draft would defer to publish, and the failure most people hit on their first task.	`no OpenAI API key for this workspace - add one under BYOK`
A tool that does not exist, or has no published version	`tool 'get_order' has no published version - publish it before attaching`
An integration that is not MCP-capable, or is not `active`	`integration 'Northwind CRM' is needs_reconnect - only active MCP integrations can be attached`
Two attached MCP servers sharing a `tools_namespace`	Names the namespace.
Two attached tools exposing the same name to the model	Names the tool.
An unknown LLM provider or model, or a `reasoning_effort` that model does not offer	Lists what the catalog accepts.
A `{{token}}` in the prompt with no root of ours — `{{name}}`, `{{customer.number}}`	`the system prompt reads {{customer.number}}, which is not a variable …`
`{{system_vars.x}}` that is not a system variable	Lists the system variables.
`{{system_vars.now}}` / `.date` / `.time` with no `timezone` set	Names the field to set.
`{{secrets.NAME}}`, `{{args.x}}` or `{{tooldata.x}}` in the prompt	Says why each one cannot be there.
The same name twice in `output`	`these output fields are declared more than once: …`
A name that is both a variable and an output field	`these names are both a variable and an output field: …`
A task name already used in the workspace	`a task with that name already exists`
An inline tool or MCP server whose name already exists (`409`)	Reference the existing one by id, or rename.
Anything the config shape forbids (`422`)	A field a task does not have — `greeting`, `kb_ids`, `channel` — an `output` list that is empty or longer than 25, an output field name outside `^[a-z][a-z0-9_]{0,39}$`, an empty field description, `max_steps` outside 1–50, `timeout_seconds` outside 10–600, a `timezone` that is not an IANA name.

DELETE /v1/tasks/{task_id} is refused with a 409 while an email batch is still drafting with that task, naming the batches. Cancel them first. Deleting a task that finished drafting a batch is fine — the batch keeps rendering, and so does the run history.

Warnings

create_task and update_task return warnings beside the saved task. These do not block the save, and there is no publish screen to read them on, so read them where they come back.

Warning	Why it matters
The prompt reads `{{vars.x}}` that the task does not declare	It resolves to nothing on every run. Usually a typo.
The task has no prompt at all	The only instruction the model gets is the output field list and the generated sentence telling it to call `submit_result`.
The prompt reads `{{system_vars.human_phone_number}}`, `.agent_phone_number` or `.direction`	There is no call, so these are always empty.

Every limit

Limit	Value
`output` fields	1 to 25
`vars` declarations	at most 25
Output field name	lower-case letters, digits and underscores, starting with a letter, up to 40 characters
Output field types	`string`, `boolean`, `integer`, `number` — no arrays, no nested objects
`max_steps`	1 to 50, default 25
`timeout_seconds`	10 to 600, default 120
Trace entries kept per run	200
Trace field length	4,000 characters, then cut and flagged `truncated`
Runs per page	default 50, maximum 200
Tasks per page	default 200, maximum 200
Versions of a task	none — one config, no history, nothing to roll back to

Get started

Agents

Tools

Knowledge

Agent teams

Agent tasks

Telephony

Web & chat

Integrations

Email outbound

Observability

Guides

Platform

Errors and limits

The two budgets

`error.type`

Refused when you run

Refused when you save

Warnings

Every limit

​The two budgets

​error.type

​Refused when you run

​Refused when you save

​Warnings

​Every limit

The two budgets

`error.type`

Refused when you run

Refused when you save

Warnings

Every limit