Agents That Teach: Towards Designing Incidental Learning Back into AI-Assisted Software Development
Citation. Mehra, R., Suri, S., Tagadinamani, P. K., Singi, K., Kaulgud, V., & Burden, A. P. (2026). Agents That Teach: Towards Designing Incidental Learning Back into AI-Assisted Software Development. arXiv:2607.06101.
Summary. This design paper argues that delegating coding tasks to AI agents can short-circuit incidental learning: developers may receive working changes without practicing the reasoning, API knowledge, debugging, and trade-off assessment that build software-engineering expertise. It names the resulting gap Knowledge Debt and proposes six principles for learning-aware developer–agent interactions. The authors present SHIELD, an early multi-agent prototype that uses coding-agent telemetry to identify, assess, and asynchronously address likely knowledge gaps without interrupting development flow.
Key claims
- Knowledge Debt is proposed as a developer-level analogue of technical debt: AI-executed changes that a developer does not fully understand can accumulate a hidden liability that becomes costly when independent debugging, adaptation, or extension is later required.
- The paper proposes six design principles for reinserting incidental learning into agentic software development: interventions should be contextual, grounded in agent reasoning, ambient, selective, adaptive to the developer, and closed-loop through comprehension checks.
- SHIELD operationalizes those principles through telemetry observation, teachability triage against a developer concept map, asynchronous probes, contextual microlearning, and post-learning assessment; it is designed to preserve developer control over when to engage.
- The paper treats productivity and learning as potentially complementary design outcomes, but it does not claim that SHIELD improves either outcome because empirical user studies remain future work.
Evidence & limitations
- This is an arXiv preprint and design/prototype paper, not an outcome evaluation; it reports no dataset or controlled user study and explicitly positions empirical evaluation of learning and productivity as future work.
- The source is publicly verifiable: the downloaded Drive PDF exactly matches the public arXiv PDF by SHA-256; the arXiv API was unavailable during ingest, so public provenance relies on the stable arXiv landing page, embedded arXiv metadata, and exact public-PDF hash match.
- The proposed use of agent reasoning telemetry and developer concept maps raises untested validity, privacy, worker-monitoring, and pedagogical-quality questions that matter before organizational deployment.