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:
- Clarify the problem and unknowns.
- Use research to fill knowledge gaps.
- Convert research into an agent-readable execution plan.
- Give agents bounded, tool-using work.
- Let runtime events reveal bottlenecks.
- Review and clean up as a separate job.
- 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 Mode | Description | Fix |
|---|---|---|
| Micromanagement | Human babysits every turn and destroys autonomy | Check in periodically, not constantly |
| Wrong problem space | Agent works on fuzzy objective with no feedback loop | Choose testable tasks |
| Context collapse | Long-running work loses rationale and handoff state | Write transition plans |
| Technical debt | Agent iterates messily and leaves clutter | Schedule cleanup as its own task |
| Model mismatch | Expensive model used for bulk work, cheap model used for hard judgment | Allocate models by task |
| Overtrust | Human skips review because output looks competent | Keep 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