hebel. today · 15 November 202633 / 79
Agentic Engineering · AI engineering
Coding agents trade re-planning for reusable workflows
Coding agents historically consume excessive tokens by re-planning every step of a task during each execution. Developers now use Swamp to convert repeated skills into persistent workflows where code manages the execution loop, invoking…
From 4 newsletters · 1 min read · one thing to try
Coding agents historically consume excessive tokens by re-planning every step of a task during each execution (Daily Dose of Data Science). Developers now use Swamp to convert repeated skills into persistent workflows where code manages the execution loop, invoking the model only for specific steps that require judgment (Daily Dose of Data Science). This architecture replaces the agent's tendency to rediscover processes, reducing token usage by approximately 8x and cutting execution time in half for certain workflows.
Tun Shwe, a staff AI engineer, defines the agent harness as a system comprising context management, tool registries, guardrails, loops, and verifiers (Packt Deep Engineering). This structure addresses intent debt and agentic amnesia, where autonomous systems lose context or drift from original goals over time (Douglas Levin from Lessons from a Startup Life). By treating memory as an append-only event log outside the model's window, engineers ensure the agent maintains a persistent, coherent state across long-running tasks.
Reliability in production requires moving beyond standard uptime metrics to sample agent outputs against defined business rules. While 78% of enterprise leaders have active agent pilots, only 14% have successfully scaled them to operational use due to these semantic failures. Teams at Uber now use a system that produces over 70% of pull requests by implementing shared rules files and human checkpoints at the task, planning, and merge stages (Ruben Dominguez from The AI Corner).
Try today · 45 min
Convert one repeated agent task into a Swamp-style typed workflow
Identify a sequence where your agent performs the same steps, then extract the logic into a script that calls the model only for judgment.
You need
- Existing agent script
- List of steps in the task
- Definition of inputs and outputs for each step
Steps
- Audit your agent's recent logs to identify a repetitive task.
- Extract the decision-making logic into a Python script.
- Replace model-driven re-planning with a hard-coded control loop.
- Isolate the specific steps requiring LLM judgment.
- Refactor those steps into independent, typed model calls.
- Verify the output against your baseline performance.