Human-orchestrated Devin fleet — 8-12x efficiency on a 6M-LoC migration.
Recent activity
Version cuts and proof, newest first — the living track record.
Spec sheet
The benchmark fields — designed for comparison across teams.
- Topology
- Orchestrator–Worker
- Agent count
- 1
- Platform
- Devin
- Runs on
- Devin
- Industries
- fintech
- Task kinds
- code-migration
- Trust tier
- Self-Reported
- Proof entries
- 1
Topology & roster
Orchestrator-worker with a human orchestrator. A human PM/engineer team prompts, reviews, and approves many parallel Devin instances working decomposed migration subtasks; there is no AI orchestrator agent in this pattern.
System wiring
Node details
Typical Orchestrator–Worker layout — schematic, not verified wiring
HumanHuman operatorHuman gate
- Tool
- Human operator
- Autonomy
- Human-gated
- directs → Orchestrator
OrchestratorOrchestrator
- Tool
- Orchestrator
- Autonomy
- Runs autonomously
- dispatches → Worker agent
- requests gate → QA reviewer
- directs ← Human operator
BuilderWorker agent
- Tool
- Worker agent
- Autonomy
- Runs autonomously
- commits to → Shared workspace
- dispatches ← Orchestrator
QAQA reviewer
- Tool
- QA reviewer
- Autonomy
- Runs autonomously
- gate verdict → Shared workspace
- requests gate ← Orchestrator
ResourceShared workspace
- Tool
- Shared workspace
- Autonomy
- Runs autonomously
- commits to ← Worker agent
- gate verdict ← QA reviewer
How a typical Orchestrator–Worker team handles a task
Typical Orchestrator–Worker layout — schematic, not verified wiring
Task arrives
Human operator directs Orchestrator.
The orchestrator routes the work
Orchestrator dispatches build work to Worker agent.
Worker agent builds the work
Worker agent builds the work.
Independent review gates the work
QA reviewer reviews the work. This reviewer is autonomous and separate from the agent that built the work, so the check is independent of its author.
The artifact lands
The artifact lands in Shared workspace: Worker agent contributes via "commits to" and QA reviewer contributes via "gate verdict".
Human holds the last word
Human operator holds final approval.
Replicate a typical Orchestrator–Worker setup
Typical Orchestrator–Worker layout — schematic, not verified wiring
Ingredients
- HumanHuman operator
- OrchestratorOrchestrator
- BuilderWorker agent
- QAQA reviewer
- ResourceShared workspace
Setup order
- 1.Provision the substrate: Shared workspace.
- 2.Stand up the orchestrator: Orchestrator.
- 3.Wire Worker agent: it receives "dispatches" from Orchestrator. Wire QA reviewer: it receives "requests gate" from Orchestrator.
- 4.Give QA reviewer an independent workspace/verdict channel: "gate verdict" to Shared workspace.
- 5.Declare the human gate: Human operator holds final approval.
Performance metrics
Windowed metrics with provenance. [unknown] means it was not tracked — an honest hole beats an invented figure.
'>20× cost savings' — cost of running Devin vs the hourly cost of an engineer on the delegated scope. This figure appears ONLY on the vendor (Cognition) page; Nubank's own engineering blog does not repeat it — rests on the vendor's telling alone, not independently corroborated by the customer. Parallel-instance count is [unknown] and not stated as a number anywhere. Source: devin.ai/customers/nubank, 2024-12. [self_reported]
'≥8× (range 8–12×)' engineering-hours efficiency gain — typical hours to complete a data-class migration task vs total hours spent prompting/reviewing Devin. Lower bound of the vendor's stated range is stored here; full range in this note. Scope: 8-year-old, 6M+ LoC monolith; ~100,000 data-class implementations; an estimated 18 months × 1,000+ engineers of work; the Data, Collections, and Risk business units each finished in weeks. Corroborated in substance by Nubank's own engineering blog (12× efficiency, weeks-not-months). Fine-tuning cut per-subtask time from ~40 to ~10 minutes — stated directly on the primary vendor page. Source: devin.ai/customers/nubank, 2024-12; building.nubank.com engineering blog, 2025-03. [self_reported]
Token economics
Cost transparency is part of the honesty architecture. [unknown] means it was not tracked — not that it is zero.
Blueprint
Operational DNA — why it works, how it was built, and how it is overseen. Not files for sale; knowledge of the design.
Fine-tuning on prior migration examples plus a fixed human-review gate lets the fleet scale horizontally (many parallel instances) without scaling the review burden proportionally — the efficiency gain comes from parallelism bounded by a cheap human checkpoint, not from removing oversight.
Cognition's Devin, task-specific fine-tuned on examples of prior Nubank migrations. Deployed as many parallel instances against decomposed subtasks of the ~100,000-data-class migration.
Human-in-the-loop by design: engineers review and approve every Devin change before it merges. This is not an autonomous pipeline — the human review step is the safety mechanism.
Proof (1)
The team's shared track record — tasks, incidents, lessons, milestones. Per-entry provenance tags are always visible.
- ArtifactDec 1, 2024self-reported
Devin @ Nubank case study published (Cognition + Nubank)
8-12x efficiency gains, >20x cost savings (vendor page only) on a 6M+ LoC ETL migration; corroborated in substance (12x, weeks-not-months) by Nubank's own engineering blog (2025-03).
https://devin.ai/customers/nubank/
Sign in to add a proof entry.
Sign inAttestations (0)
Named third-party statements from people with first-hand experience. Attestations are what separates Peer-Attested from Evidence-Linked.
No attestations yet. Worked with this configuration or agent? Attest to it using the form below — attestations are named third-party statements and are what separates Peer-Attested from Evidence-Linked.
Sign in to attest to this team.
Sign in