Source: https://internal-agents.com/agents/zup-codegen # Zup — CodeGen CodeGen is Zup's internal coding agent, described in a paper by its builders. They report that targeted tool design, such as string-replacement edits over full-file rewrites, and layered safety guardrails mattered more than prompt engineering. Progressive human oversight drove adoption. - Approach type: Task agent - Deployment stage: Research - Autonomy: Human in loop - Evidence strength: Detailed primary - Status: Internal - First reported year: 2026 - Work: Coding - Invocation: Unknown - Entry reviewed: 2026-09-15 ## Where people stay involved - **constrained coding task → human-supervised edit** — Continuous steering · Level 2 ## Overview ### Summary CodeGen is Zup's internal coding agent, described in a paper by its builders. They report that targeted tool design, such as string-replacement edits over full-file rewrites, and layered safety guardrails mattered more than prompt engineering. Progressive human oversight drove adoption. Fact · Reported · High confidence · `zup-codegen--summary` Confidence reason: A linked first-party source states the claim. Evidence: - Supports · [1] [Building an Internal Coding Agent at Zup](https://arxiv.org/abs/2604.09805) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/zup-codegen-source-1/content.md) ## How it works ### Supporting component Tools that limit what the agent can change, vs. unconstrained generation Fact · Reported · High confidence · `zup-codegen--primitives-0` Confidence reason: A linked first-party source states the claim. Evidence: - Supports · [1] [Building an Internal Coding Agent at Zup](https://arxiv.org/abs/2604.09805) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/zup-codegen-source-1/content.md) ### Supporting component Levels of human review that increase trust over time Fact · Reported · High confidence · `zup-codegen--primitives-1` Confidence reason: A linked first-party source states the claim. Evidence: - Supports · [1] [Building an Internal Coding Agent at Zup](https://arxiv.org/abs/2604.09805) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/zup-codegen-source-1/content.md) ## Supervision evidence ### Operating model assessment Level 2 for constrained coding task → human-supervised edit; human attention boundary: continuous-steering. Inference · Catalog judgment · Medium confidence · `zup-codegen--operating-models-0` Confidence reason: The research source describes progressive human oversight but does not establish background delegation. Qualifications: - Observation date: 2026 Evidence: - Supports · [1] [Building an Internal Coding Agent at Zup](https://arxiv.org/abs/2604.09805) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/zup-codegen-source-1/content.md) ## Implementation details ### Harness Constrained editing tools; state-management concerns; progressive levels of human oversight Fact · Reported · High confidence · `zup-codegen--architecture-harness` Confidence reason: A linked first-party source states the claim. Evidence: - Supports · [1] [Building an Internal Coding Agent at Zup](https://arxiv.org/abs/2604.09805) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/zup-codegen-source-1/content.md) ### Model Not specified (see paper) Fact · Reported · High confidence · `zup-codegen--architecture-model` Confidence reason: A linked first-party source states the claim. Evidence: - Supports · [1] [Building an Internal Coding Agent at Zup](https://arxiv.org/abs/2604.09805) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/zup-codegen-source-1/content.md) ### Tool access Constrained editing tools rather than free-form code generation Fact · Reported · High confidence · `zup-codegen--architecture-tool-access` Confidence reason: A linked first-party source states the claim. Evidence: - Supports · [1] [Building an Internal Coding Agent at Zup](https://arxiv.org/abs/2604.09805) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/zup-codegen-source-1/content.md) ### Credentials Layered safety controls Fact · Reported · High confidence · `zup-codegen--architecture-credentials` Confidence reason: A linked first-party source states the claim. Evidence: - Supports · [1] [Building an Internal Coding Agent at Zup](https://arxiv.org/abs/2604.09805) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/zup-codegen-source-1/content.md) ## Lessons and interpretation ### Lesson Targeted tool design and layered safety controls matter more than prompt tweaks Inference · Catalog judgment · Medium confidence · `zup-codegen--lessons-learned-0` Confidence reason: The catalog derives this observation from the linked sources. Evidence: - Supports · [1] [Building an Internal Coding Agent at Zup](https://arxiv.org/abs/2604.09805) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/zup-codegen-source-1/content.md) ### Lesson Progressive levels of human oversight help build trust before granting autonomy Inference · Catalog judgment · Medium confidence · `zup-codegen--lessons-learned-1` Confidence reason: The catalog derives this observation from the linked sources. Evidence: - Supports · [1] [Building an Internal Coding Agent at Zup](https://arxiv.org/abs/2604.09805) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/zup-codegen-source-1/content.md) ## Sources 1. [Building an Internal Coding Agent at Zup](https://arxiv.org/abs/2604.09805) - Paper · First party · Evidence - Original URL: - Accessed: 2026-08-12 · Last verified: 2026-08-31 - Preserved copy in the repository: