DeepLearning.AI's new course: three adaptation layers so coding agents stop relearning every session
DeepLearningAI · youtube · 2026-08-26
DeepLearning.AI partnered with Oracle on a short course, Building Adaptive AI Agents, taught by Oracle's Nacho Martínez and Casius Lee. Core idea: today's coding agents start from zero every session — an environment issue fixed in minutes yesterday gets rediscovered, burning the same time and tokens today.
The course covers three adaptation layers:
- Behavior adaptation: build a skill induction pipeline that turns agent traces (conversations, tool calls, errors, fixes) into reusable skills retrieved on similar tasks, with a human approving each skill.
- Knowledge adaptation: construct a code knowledge graph from imports, function calls and co-edits in git history, so the agent retrieves context by navigating the codebase structure instead of keyword search — the course argues retrieval, not code generation, is often the real bottleneck in large codebases.
- Model adaptation: when weight-space adaptation like fine-tuning is the right tool.
By the end you can build agents that get measurably better at the same work day over day instead of paying twice for the same mistakes.
Related event: DeepLearning.AI Launches Course on Adaptive Coding Agents(2 posts)→
More from coding & agent
- Cloud Agents vs. Local Machines: Trust and Verification Workflows — hugobowne · 2026-08-27
- Deepagents Evolving into Multiplayer Harness — hwchase17 · 2026-08-27
- Google proposes upgrading OKF to enterprise infrastructure with Knowledge Catalog — gaganghotra_ · 2026-08-27
- WebMCP May Become Standard for Future Sites, Accelerating Agent Interaction — thisiskp_ · 2026-08-27
- Cloudflare open-sources security-audit-skill: parallel agents run six-phase vulnerability audits — tom_doerr · 2026-08-27
- ARC-AGI-3 learnings: Key designs for self-learning AI agents — GregKamradt · 2026-08-27