Atlas task fails with the subagent's own error text, or with Task cancelled, because TaskTool inspects the finished job status: an error status re-throws the child's error message, falling back to "Task failed", and a cancelled status throws "Task cancelled". The fix is to read the child's error text in the message, since it is passed through verbatim, then open the subagent's session to inspect its transcript rather than guessing from the summary line.
Why does Atlas report Task failed or Task cancelled
Atlas reports Task failed or Task cancelled because TaskTool inspects the finished job status of the subagent. In the 2026 source, an error status re-throws the child's error message and falls back to "Task failed" when there is none, while a cancelled status throws "Task cancelled".
The two messages describe different endings, and telling them apart is the first diagnostic step. Task cancelled means the job was stopped, not that it broke. A subagent error, by contrast, arrives as the child's own error text, because TaskTool re-throws it rather than replacing it with a generic string. The literal words "Task failed" only appear as a fallback, when the child errored without producing a message of its own. So a bare "Task failed" is itself information: the subagent failed and said nothing useful about why, which is your cue to go read its transcript.
How to read the real error from an Atlas subagent
Read the child's error text in the Atlas message. TaskTool passes it through verbatim, so the string you see in the parent session is the subagent's own error, not a paraphrase. The generic Task failed appears in exactly 1 situation, as a fallback when the child produced no message of its own.
Treat the surfaced text as the primary evidence. If the child failed on a tool call, a permission denial, or a bad path, that error is what you are looking at, and it usually names the tool and the offending argument directly. Because every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, a subagent can fail on a rule that the parent session never encounters, which is one of the more common surprises. Match the error text to the tool it came from before you change anything about the task prompt.
How to open the subagent session transcript in Atlas
Open the Atlas subagent's session to inspect its transcript rather than guessing from the summary line. The parent session shows 1 line, either the child's error text or Task cancelled, while the child's session holds the full sequence of tool calls and model turns that led there.
The summary line is a verdict, not an explanation. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and each of those subagents keeps its own transcript. Read it end to end: you will see which files the child opened, which searches it ran, where it lost the thread, and which tool call produced the failure. Guessing from the one-line summary is how developers end up rewriting a prompt that was never the problem, when the actual failure was a missing language server or an unreachable endpoint several steps in.
Why did my Atlas subagent get cancelled
An Atlas subagent shows Task cancelled when its finished job status is cancelled rather than errored, and TaskTool treats those 2 statuses differently. If you cancelled the parent session, the child is cancelled too, and that is expected behavior, not a bug in the task tool.
Cancellation propagates downward. Interrupting the parent stops the children it launched, because a subagent exists to serve the parent's turn and has nothing to return to once that turn is gone. When you see Task cancelled and you remember hitting the interrupt, the loop is closed and there is nothing to debug. When you see Task cancelled and you did not cancel anything, look at what else stopped the session, since the child did not error on its own. The distinction is worth checking before you spend time on a prompt rewrite.
How to re-run a failed Atlas task successfully
Re-run the Atlas task with a narrower prompt if the child ran out of context or tools. 1 subagent handed a scope it cannot finish will exhaust its budget and end with an error status, which TaskTool then re-throws into the parent session as the child's own error text.
Narrowing works in two directions. Reduce the surface area of the question, asking about one module rather than an entire service, and reduce the tool surface the child needs, pointing it at files you have already located instead of asking it to find them from scratch. Both shrink the work per subagent. If several narrow tasks would cover the ground, run them: Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, so several small tasks are usually a better shape than one that keeps failing.
How to fix it
- 01Read the child's error text in the message; it is passed through verbatim.
- 02Open the subagent's session to inspect its transcript rather than guessing from the summary line.
- 03If you cancelled the parent session, the child is cancelled too; that is expected.
- 04Re-run the task with a narrower prompt if the child ran out of context or tools.
Frequently asked questions
- how to debug Task failed in Atlas
- Read the child's error text in the message, because TaskTool passes it through verbatim, then open the subagent's session to inspect its transcript rather than guessing from the summary line.
- what does Task cancelled mean in Atlas
- Task cancelled means the subagent's job status was cancelled rather than errored. If you cancelled the parent session, the child is cancelled too, and that is expected.
- why does Atlas show a generic Task failed with no details
- "Task failed" is the fallback TaskTool uses when the child errored without producing an error message of its own. Open the subagent's session transcript to see what actually went wrong.
- does cancelling an Atlas session cancel its subagents
- Yes. If you cancelled the parent session, the child is cancelled too. Cancellation propagates from parent to subagent, and Task cancelled is the expected result.
- how do I fix an Atlas subagent that runs out of context
- Re-run the task with a narrower prompt. Reduce the scope of the question and point the subagent at files you have already located instead of asking it to search from scratch.
- is the Atlas subagent error message the real error
- Yes. TaskTool re-throws the child's error message verbatim, so the text in the parent session is the subagent's own error, not a rewritten summary.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas with Kimi K2 Turbo: Pricing, Context, and Setup in 2026
Kimi K2 Turbo in Atlas costs $2.40 per Mtok input and $10.00 per Mtok output on a 256K tokens (262,144) window. A latency purchase, not a capability upgrade.
Atlas with Grok Build 0.1: xAI's First Coding-Agent Model (2026)
Grok Build 0.1 runs in Atlas at $1 / $2 per Mtok with a 256K context and 256,000 max output tokens. A coding-specialized model, but it is a 0.1 release.
Audit a Repo with Parallel Subagents in Atlas (2026 Workflow)
How to audit a repo with parallel subagents in Atlas in 2026: the task tool launches explore subagents in their own sessions, so only conclusions return to your context.
Atlas with Gemini 3 Flash (2026): The $0.5 Default for All Day Agent Loops
Gemini 3 Flash gives Atlas a 1,048,576 token context and 65,536 token output at $0.5 per Mtok input and $3 per Mtok output. The economical default for agent loops.
Atlas with Qwen2.5 72B Instruct: The Flagship Dense Qwen in 2026
Qwen2.5 72B Instruct in Atlas: 128K tokens (131,072), $1.40 per Mtok input, $5.60 per Mtok output, openly published weights you can serve on your own vLLM.
Atlas for Actix Web in 2026
Atlas is a terminal-native AI coding agent for Actix Web in 2026. It reads extractors and app_data, then runs cargo test and cargo clippy behind a prompt.
Run the Test Suite and Triage the Failures with Atlas in 2026
How to triage a failing test suite with Atlas in 2026: bash truncates at 2000 lines or 50 KB and saves the full log, then grep groups failures by root cause.
Atlas with Llama 3.1 8B (Ollama): 128K Context on an 8GB Card in 2026
Run Atlas on Llama 3.1 8B (Ollama): Meta's 4.9GB workhorse with a 128K context, free self-hosted. Setup, the KV cache catch, and when a coder model wins.