Advanced Agentic Engineering

Five first principles for scaling agentic development -- alignment, feedback loops, chaotic oscillation, context contamination, and the review bottleneck.

This itinerary is for engineers and tech leads who are already building with AI agents and want to reason from first principles about how to scale agentic development across teams and codebases.

The central thesis: agentic engineering is an alignment problem. Alignment means how close the agent’s actions and outputs stay to human expectations and intent. Every design decision becomes a question of alignment as output scales beyond what any human can review line by line.

Principle I: Agentic Time >>> Human Time

Agentic code cost tends to zero. Human attention does not. The optimal agentic workflow may look wildly different from traditional software engineering.

When an agent stalls or produces partial output, don’t step in — push the agent to continue with accumulated context.

  1. What Are Agents?
  2. Workflows vs. Agents
  3. Ralph Technique — the push loop that combats lazy completion
  4. Orchestrator-Worker

Principle II: Misalignment & Feedback Loops

Intelligence is contextual: an agent without tooling is fragile, an agent with linting, testing, and type checking is self-correcting. But not everything is deterministically testable. Give agents an explicit escape hatch to pause and ask.

  1. Human-in-the-Loop — the HITL tool that changes the action space
  2. Evaluator-Optimizer
  3. Reflection

Principle III: Chaotic Oscillators

Agentic errors compound through context contamination and refactor avoidance. Agents find shortcuts that satisfy tests while missing intent. Monitor the reasoning process, not just the output.

  1. Reward Hacking — the central failure mode
  2. Hawk Agent — cross-provider process-level oversight
  3. Parallelization

Principle IV: Memetic Virus

Context is not neutral. Bad patterns in the codebase propagate through every agent session that touches them. When a direction fails, don’t fix forward — extract learnings, rollback, restart clean.

  1. Rollback and Learnings — recovery when the cost of failure approaches zero
  2. Iterative Refinement
  3. Plan-and-Execute

Principle V: The Review Bottleneck

Agent output grows non-stop; reviewer time stays constant. Pull requests as designed will not survive agentic scale. New patterns: surprise-driven triage, artifact-based review, agent timelines. New teams: stronger ownership, agent-first interfaces, faster infrastructure.

  1. Agentic Code Review — review patterns for agentic scale
  2. Agent-First Teams — how organizations restructure
  3. Handoffs
  4. Composite Patterns

Outcome

At the end of this itinerary you should be able to design an agentic engineering workflow for a team: feedback loops, oversight patterns, review processes, and the organizational changes required to make it work at scale.