Internal Agents Map

CodeGen

CodeGen is Zup's internal coding agent. In the preserved paper abstract, its authors attribute reliability improvements to editing tools and layered controls, and adoption to progressive human oversight. The abstract provides no measurements or definitions of those oversight modes.

Company
Zup
Approach type
Agent
Work
Coding
Human involvement
Human in loop
Deployment stage
Research
Evidence strength
Detailed primary
Entry reviewed
Copy as MDAsk ChatGPT

How it works

Not reported: The abstract reports design decisions and outcomes, not a run; no trigger-to-output sequence is described.

Where people stay involved

People review, approve, or guide the agent’s work here.

Reported: Progressive human oversight modes are reported as present; the abstract does not define their boundaries.

  • Supervision level

    Not classified

    • constrained coding task → human-supervised edit[1]

How supervision levels work

Implementation details

Model
Not reportedThe abstract names no model and argues that the surrounding engineering matters more.
Harness
Constrained editing tools; state-management concerns; progressive levels of human oversight[1]
Sandbox
Not reported
Tool access
Constrained editing tools rather than free-form code generation[1]
Knowledge
Not reported
Context management
Not reportedThe abstract names state management as a decisive concern without describing any management mechanism.
Credentials
Not reportedThe paper names layered safety guardrails, which are safety enforcement, not credential handling.
Interfaces
Not reportedNo interface is named in the abstract.

Mechanisms

Constrained editing tools

String-replacement edits instead of full-file rewrites, combined with layered safety controls[1]

Progressive oversight

Progressive oversight modes reported by the authors; the preserved abstract does not define the boundaries[1]

Validation and failure handling

Not reported: Layered safety guardrails are named as a design element; no check of the agent's work products is documented.

Reported observations

Not reported: The abstract states qualitative conclusions without measurements; the record carries no observation claims.

Lessons

Reported opinion: Zup's authors report larger reliability gains from targeted editing tools and layered controls than from prompt changes in CodeGen.[1]

Reported opinion: Zup's authors attribute organic adoption partly to progressive human oversight modes.[1]

Back to agents
Improve this record

Give your coding agent the context and instructions to prepare a source-backed update.

Suggest on GitHub

Paste the prompt into your coding agent, then describe the correction or add a public source.

Preview prompt

Sources and research details

Citations link to the original publisher. Each source also keeps a preserved copy in the repository, so a changed or removed page stays checkable.

  1. Building an Internal Coding Agent at Zuphttps://arxiv.org/abs/2604.09805Paper · First party · Last source verification: 2026-08-31
    Preserved copy
Question coverage and scope
purpose
Reported
workflow
Not reported: The abstract reports design decisions and outcomes, not a run; no trigger-to-output sequence is described.
human involvement
Reported: Progressive human oversight modes are reported as present; the abstract does not define their boundaries.
implementation
Reported
validation
Not reported: Layered safety guardrails are named as a design element; no check of the agent's work products is documented.
observations
Not reported: The abstract states qualitative conclusions without measurements; the record carries no observation claims.
lessons
Reported
Research details for every claim on this page
  1. Summary
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  2. Harness
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  3. Tool access
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  4. Constrained editing tools
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  5. Progressive oversight
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  6. Lesson
    Statement type
    Opinion
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    The preserved abstract makes that comparison and names string-replacement edits, but contains no experiment details or numerical results to assess it independently.
  7. Lesson
    Statement type
    Opinion
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    The abstract states this conclusion without specifying the modes, their review boundaries, or an adoption measurement method.
  8. Operating model assessment
    Statement type
    Inference
    Provenance
    Catalog judgment
    Confidence
    Unverified
    Confidence reason
    The abstract names progressive human oversight modes but no passage establishes the oversight form: continuous steering, per-run work-product review, and outcome review are all consistent with it.
    Observation date
    2026