An internal agent system on Amazon Bedrock with agent identities, assigned managers, and evaluation scores. Atlas prepares feature changes; Morphex can merge and ship changes that pass the required checks without human review.
Sphera agent feature task from an inbox trigger to production ship; Morphex merges autonomously above the confidence threshold while Atlas requests human review after the sandbox gate.
Receive a task from one of three inboxes
A Slack @mention, monday item assignment, or GitHub PR review request lands in the same agent session through SNS and a per-team SQS queue[Source 1]
Implement the feature task
The agent picks up tickets, writes the PR, and ships the feature from the same backlog as human teammates[Source 1]
Auto-deploy the PR to a remote sandbox
Every PR auto-deploys to a per-session remote sandbox; tests, checks, and replayed production traffic run before any human review[Source 1]
Merge above the confidence threshold
PRs above the confidence-score threshold merge automatically after passing every gate and ship to production with no human in the loop; below-threshold PRs route to a human with the failing signal named[Source 1]
Where people stay involved
People review, approve, or guide the agent’s work here.
Reported: Morphex runs exception-only: a below-threshold PR routes to a human and an approved PR ships with none. Atlas runs sandbox checks before asking for human review; the source calls threshold auto-merge a following-two-quarters goal for that population.
Supervision level
Exception-only
A person returns when the system raises an exception.
5/5Level 5 of 5
Morphex feature task → PR passing CI, remote sandbox, and Guardrails → automatic merge and production ship[Source 1]
Reported: The four run steps describe the shared Sphera pipeline; the final gate differs by agent as the scope states.
human involvement
Reported: Morphex runs exception-only: a below-threshold PR routes to a human and an approved PR ships with none. Atlas runs sandbox checks before asking for human review; the source calls threshold auto-merge a following-two-quarters goal for that population.
implementation
Reported
validation
Reported: Validation is the blocking Guardrails reviewers, the remote-sandbox tests with replayed production traffic, and the deterministic and LLM-scored evals fed back into the harness.
Morphex: 19 of 20 PRs merge automatically without human review
Duplicate representation of Headline claim. Same nineteen-in-twenty automatic-merge figure as the headline with the same denominator and unset period; the word automatically adds nothing the headline scope does not already state.
Statement type
Metric
Provenance
Reported
Confidence
Not assessed
Confidence reason
No explicit confidence assessment is recorded for this claim.
Reported by
monday.com
Scope
Morphex PRs that merge automatically after CI and Guardrails pass
Denominator
Morphex pull requests; sample size and period not supplied
The authors describe replacing a looks-fine standard with deterministic metrics and model-scored evaluations, and explicitly regret waiting until month nine.
The memory section describes failed context stuffing and vector retrieval for Atlas, then names the file lifecycle. It does not compare retrieval methods on other workloads.
The article describes the evaluation feedback and Guardrails reviewers separately; these mechanisms do not guarantee that every accepted change is safe.
The source states that nineteen of every twenty Morphex PRs merge automatically by passing every gate and that approved PRs ship to production with no human in the loop; a PR below the confidence threshold routes to a human.
The evidence is two-sided for Atlas; the source describes tests running before a request for human review while also describing confidence-threshold auto-merge as being driven out over the following two quarters.
