1 Minute Signal

"My agentic engineering workflow" | George Pickett (MTS @ Parallel)

Video thumbnail: "My agentic engineering workflow" | George Pickett (MTS @ Parallel)
Jun 30, 202624m 6s video lengthGreg Kamradt

The Signal

George, a software engineer who builds autonomous coding agent workflows, argues that the central failure point in long-running AI tasks is the loss of intent across compactions. He contends that rather than relying solely on goal-based execution, practitioners should maintain human oversight via an initial 'grill' interview and durable markdown artifacts that keep the agent aligned with the original mission. The core dispute remains whether these complex chains of decisions and executable plans are truly necessary for success or if autonomous goal execution with occasional steering can reach the same results.

The Case

  • George uses an agent-led interview protocol called "grill me" to extract specific user intent before any code is written, a phase he identifies as the highest-leverage investment of human time in the development cycle.8:08
  • To prevent context drift, he requires agents to write decisions into a "decisions.mmd" file and an "exec plan," which serve as durable, gitignored reference documentation the agent can consult during long-running sessions.2:59
  • While he automates implementation through "goalcraft" and chained skill pipelines, he admits self-review often lacks critical distance and has found that external tools like Codto—a pull-request review model—consistently catch bugs his agents missed during their own self-critique.18:47
  • He stores these iterative execution plans in a repository folder labeled "agentwork," though he acknowledges these files are throwaway documents optimized for agent retrieval rather than human readability.23:22
  • George measures skill usefulness using subjective human vibes and bug-count discovery, admitting he lacks a more formal or automated benchmarking process for his current workflow.21:53
  • His claims regarding the efficiency of his grilling protocol are self-reported, with George estimating that 90–95% of the time, the agent's follow-up questions accurately capture his requirements.9:43

The 1 Minute Signal Take

This video is a useful look at the mechanics of stateful agent construction for those already familiar with the basics of LLM-based coding tasks. It avoids the fluff of standard AI demos and presents a concrete, if unverified, modular philosophy for keeping agents on track. Watch it if you want the specific prompting patterns for durable agent memory; skip it if you are looking for settled engineering benchmarks, as George's stack is largely anecdotal.

Pro Analysis

Strategic Significance: This workflow addresses the 'memory bottleneck' in long-running AI agents, moving the paradigm from reactive prompt engineering to structured, artifact-driven autonomy. It effectively treats the agent as a junior dev requiring a clear spec and frequent oversight.

Who Should Care: Software engineers and technical leads who are building agentic systems or automating complex workflows and struggle with agents failing to reach project completion without human intervention.

Contrarian Takeaway: The most effective coding agents are not the ones with the massive context windows, but the ones built with the most rigid, human-enforced constraints and externalized decision-making logs.

Time saved:21m 46s

Share this

Tags

AI agentsLLMsSoftware developmentWorkflow automation