Agentic Coding Workflows

Agentic coding workflows use AI systems as semi-autonomous execution agents rather than passive autocomplete tools. The human's role shifts from writing every line to designing the system of work: research, task boundaries, quality checks, handoffs, and cleanup.


Core Pattern

The workflow is:

  1. Clarify the problem and unknowns.
  2. Use research to fill knowledge gaps.
  3. Convert research into an agent-readable execution plan.
  4. Give agents bounded, tool-using work.
  5. Let runtime events reveal bottlenecks.
  6. Review and clean up as a separate job.
  7. Preserve context in durable notes, docs, or handoffs.

This is close to managing a small technical team, except the "team" is made of agents with uneven judgment and high execution stamina.


Good Problem Spaces

Agentic coding works best when:

  • outputs are testable
  • benchmarks are repeatable
  • the agent can run tools
  • the task is backend, infrastructure, or data-heavy
  • failures can be detected mechanically
  • design taste is less central than correctness

It works worse when:

  • the goal is fuzzy
  • user reaction is the main metric
  • inputs are scarce or unpredictable
  • the task needs strong taste, nuance, or original creative judgment

Human Role

The human becomes a bottleneck manager:

  • define the work
  • choose the model/tool for the job
  • set quality bars
  • check progress at intervals
  • remove blockers
  • force research when assumptions get stale
  • review security and technical debt
  • decide what deserves to exist

This connects to something-is-different-about-2026: AI abstracts skills upward. The scarce skill becomes judgment, taste, problem selection, and context management.


Failure Modes

Failure ModeDescriptionFix
MicromanagementHuman babysits every turn and destroys autonomyCheck in periodically, not constantly
Wrong problem spaceAgent works on fuzzy objective with no feedback loopChoose testable tasks
Context collapseLong-running work loses rationale and handoff stateWrite transition plans
Technical debtAgent iterates messily and leaves clutterSchedule cleanup as its own task
Model mismatchExpensive model used for bulk work, cheap model used for hard judgmentAllocate models by task
OvertrustHuman skips review because output looks competentKeep review and security gates

Relation To The LLM Wiki

This Second Brain is itself an agentic workflow. The human curates sources and asks questions; the agent ingests, links, summarizes, and maintains structure. The same rules apply:

  • keep raw sources immutable
  • write durable artifacts
  • update indexes and logs
  • lint periodically
  • avoid unreviewed sprawl

Sources