Comparison
SWE-Agent: Research benchmark agent vs. production overnight desk.
SWE-Agent is a research artefact — a benchmark harness that scores language-model agents on real GitHub issues. Driftlock is a production deployment that turns overnight engineering signals into signed, audit-linked pull requests on the customer’s own protected branches. The comparison here is not about solving the same problem; it is about posture. SWE-Agent optimises for benchmark coverage inside a sandbox; Driftlock optimises for governance, replay, and human approval inside the customer’s own stack.
Posture
Driftlock runs self-hosted in your VPC; no third-party LLM is ever called at the test step.
Identity
Commits sign with your cosign key in your KMS or HSM. Keys rotate on your schedule.
Audit
Seven per-PR columns — alert id, sandbox run hash, SBOM hash, approver chain, signature, diff bytes, triage verdict — replayable on your storage tier.
What they do
SWE-Agent in one paragraph.
SWE-Agent is a research benchmark harness from Princeton’s NLP group that scores language-model agents on real GitHub issues; the unit of work is one issue selected from a fixed benchmark set, and the agent runs inside the benchmark’s research sandbox.
Where Driftlock differs
Eight dimensions where the posture diverges.
Each row names one axis — autonomy, hosting, audit, signing, SBOM, sandbox, monorepo scale, CI touch — and states both sides on that axis. No feature checklist, no marketing re-framing.
Axis · autonomy scope
Them
Solves one GitHub issue selected from a fixed benchmark set; scope is bounded by the benchmark, not by a customer signal.
Driftlock
Acts only on overnight signals already in your stack (CI failure, Dependabot alert, CVE, on-call page); no goal, no run.
Axis · self-hosting
Them
Runs inside the benchmark’s research sandbox; deployment is not the artefact under test.
Driftlock
Single Helm + Terraform deployment inside your VPC, with no third-party LLM calls at any step.
Axis · audit trail
Them
Logs to the benchmark harness; the records exist to score a model, not to be replayed by a regulator.
Driftlock
Records seven per-PR columns (alert id, sandbox run hash, SBOM hash, approver chain, signature, diff bytes, triage verdict) replayable on your storage tier.
Axis · commit-signing
Them
Uses the benchmark’s identity; there is no production-grade signing flow.
Driftlock
Commits sign with your cosign key kept in your KMS or HSM; key rotation happens on your schedule.
Axis · SBOM
Them
Does not emit an SBOM as a benchmark output; SBOMs are not part of the scoring rubric.
Driftlock
Emits an SBOM hash per PR, recorded alongside the diff and the signature so reviewers can diff two SBOMs without re-running the build.
Axis · sandbox execution
Them
Runs inside the benchmark’s research sandbox; the customer does not pin the toolchain or share the team’s test cache.
Driftlock
Runs inside a hermetic sandbox that mounts the existing monorepo’s toolchain and test cache, and replays the original regression before opening the PR.
Axis · monorepo scale
Them
Evaluates single repositories from a benchmark set; large monorepos are not the design target.
Driftlock
Built for one large monorepo per deployment; the sandbox mounts the monorepo’s existing caches so a CI rerun is bounded by what CI itself already proves.
Axis · CI touch
Them
May invoke the repository’s test suite as a scoring step but does not read CI failures as a customer trigger.
Driftlock
Treats CI as the source of truth — the agent reads CI failures as triggers and reruns the original regression inside the sandbox before opening the PR.
8 axes · posture-level · no feature checklist
FAQ
Questions comparison-stage buyers ask.
See it on your stack
Compare on a real monorepo, not a slide deck.
The architecture page names the four signal sources the watcher reads; the governance page enumerates the seven per-PR columns a regulator replays. Run the comparison against one of your real monorepos on a ten-day scoped trial — you'll see the diffs Driftlock ships and the audit trail your reviewers sign.