This series documents an enterprise workflow design for working with AI coding agents.
It is not a prompt collection and it is not tied to a single tool. It is a structural approach based on observed failure modes in production environments and the gaps that exist in current agentic tooling.
Agentic coding increases throughput. It also increases the probability of shipping something we did not intend to ship. The workflows described here are designed to close that gap.
If you are new to the series, this post provides the reading order and context.
1. Failure modes: where agents fail and where we fail
- Designing Agentic Workflows: Where Agents Fail, and Where We Fail https://dev.to/danielbutlerirl/designing-agentic-workflows-where-agents-fail-and-where-we-fail-4a95
This post defines the predictable failure surfaces that emerge under volume:
- Baby-counting (silent requirement loss)
- Cardboard muffin (plausible but hollow implementations)
- Half-assing (working but structurally poor code)
- Litterbug (residue accumulation)
- Rubber-stamping, review fatigue, intent drift, and decision delegation
The core argument: the agent’s optimisation behaviour and human cognitive limits combine to produce structural risk. Green CI and plausible diffs are observable signals, not guarantees of correctness.
Start here. It defines the problem space.
2. From diagnosis to design
- Designing Agentic Workflows: A Practical Example https://dev.to/danielbutlerirl/designing-agentic-workflows-a-practical-example-291j
This post shifts from analysing failure to designing around it.
It introduces the foundational constraint behind the workflow:
Verification must be independent of the language model.
The post explains why proposal and verification must be separated, why work must be bounded, and why durable intent must exist outside conversational context.
This is the conceptual shape of the workflow.
3. The core loop
- Designing Agentic Workflows: The Core Loop https://dev.to/danielbutlerirl/designing-agentic-workflows-the-core-loop-166d
This post describes the operational sequence used in practice:
- Define verification gates
- Plan one bounded task
- Implement exactly one task
- Repeat until all gates pass
- Run cleanup
It explains why sessions are treated as disposable, why durable state must live outside chat history, and why commit-sized changes preserve review integrity.
This is the backbone of the methodology.
4. Supplementary commands and pressure valves
- Designing Agentic Workflows: Supplementary Commands and Pressure Valves https://dev.to/danielbutlerirl/designing-agentic-workflows-supplementary-commands-and-pressure-valves-l51
The core loop handles most work.
This post covers additional structural controls used when:
- The issue cannot yet be defined in verifiable terms (
/wf-investigate) - Interfaces must be explicit before implementation (
/wf-design) - Architectural decisions introduce durable constraints (
/wf-adr) - Long sessions begin degrading context (
/wf-summarise)
These are not alternative workflows. They are pressure valves that keep the core loop viable under complexity and long-running work.
What this series is
This is a workflow design for enterprise environments where:
- Reviewability matters
- Audit and compliance exist
- Availability and security have real cost
- Silent regressions are unacceptable
It addresses gaps present in today’s tooling by changing the environment the model operates in rather than attempting to make the model smarter.
As agentic tooling evolves, some of these controls may become redundant. Today, workflow design remains the most reliable control surface available.
If you are adopting agentic coding in production systems, read the failure modes first. Then layer structure deliberately.
Throughput is visible. Regressions often are not.
This series is about narrowing that gap.
Top comments (0)