hebel. today · 9 November 202610 / 74
Agentic Engineering · AI Agent Management
Engineering operational guardrails for autonomous agents
Establishing explicit boundaries. Docusign engineering mandates that autonomous agents cannot merge pull requests or execute irreversible actions without human approval. To prevent agents from circumventing operational limits, operators must define an 'Ask-Before-Act' list of…
From 6 newsletters · 1 min read · one thing to try
Establishing explicit boundaries. Docusign engineering mandates that autonomous agents cannot merge pull requests or execute irreversible actions without human approval (Nick Lucchesi from The New Stack). To prevent agents from circumventing operational limits, operators must define an 'Ask-Before-Act' list of actions that always require human confirmation (Smart Prompts For AI). Liz Elliott, who manages 33 active brands, enforces this by maintaining a written brief that explicitly names the three categories of actions—sending emails to clients, spending money, and deleting files—that agents are strictly forbidden from performing autonomously (Liz on the Web).
Implementing durable execution. When agents run for extended periods, they frequently encounter server timeouts or connection drops that crash standard process loops (Louis-François Bouchard from Making AI Accessible with Louis-François). To mitigate this, engineering teams are shifting toward durable execution, where tasks are broken into discrete steps and results are saved externally, allowing the agent to resume from the last successful state after a failure. This approach moves beyond simple monitoring, as seen in Docusign’s practice of tracking run observability and periodic quality checks to ensure agents do not drift from their intended behavior.
Standardizing agent governance. Data lineage and metadata are increasingly vital as agents begin to read and interpret warehouse data directly (Javier Orraca-Deatcu from Agentic Bytes). To maintain consistency, Atlan has open-sourced an AI-registry that treats agent skills as versioned, owned, and fingerprinted assets, preventing the common issue of teams relying on disparate, unverified instruction files. This governance framework mirrors the 'skill' approach where professionals supply the judgment and evidence, which is then codified into reusable instructions that remain independent of any specific model or platform (Salvador Carranza from What AI makes PossibNow).
Try today · 45 min
Create a formal 'Skill' file for your most critical autonomous agent to standardize its decision-making.
Start by identifying one recurring task where you recently had to correct the agent's output.
You need
- One recurring assignment
- Source documents used for the task
- A past AI result
- Your manual correction
Steps
- Review the past AI result and your manual correction to identify the gap in reasoning.
- Draft a decision rule that explains why the correction was necessary based on the evidence.
- Format the rule into a SKILL.md file containing the task description, inputs, and decision steps.
- Create a references/examples.md file with your reviewed example and the reason for the correction.
- Link the example file to your SKILL.md and update the agent's instructions with these files.