hebel. today · 13 November 202629 / 80

Agentic Engineering · AI Agent Management

Autonomous agents require deterministic control and intent-based security

To secure production agents, Uber Engineering enforces policy decisions through an AI Registry that hooks into tool calls, denying unauthorized actions even when credentials are valid. This intent-based control limits agent autonomy to declared…

To secure production agents, Uber Engineering enforces policy decisions through an AI Registry that hooks into tool calls, denying unauthorized actions even when credentials are valid (Resilient Cyber). This intent-based control limits agent autonomy to declared constraints, countering the tendency of models to expand authority mid-task as identified by the Cloud Security Alliance (Rod Trent from Rod’s Blog). For complex workflows, Prabhaharan NM at AAIF Chennai Connect uses plain code to select tools and their sequences, calling the language model only to explain failures, which ensures that deterministic logic governs the execution path (AAIF User Community). Ioannis Zempekakis at Prosus reduced agent costs tenfold by routing tasks to smaller, fine-tuned models rather than sending every request to the most expensive model available (AAIF User Community).

  • IPI Bench found that indirect prompt injection success rates against workplace agents reached nearly 100% when attackers could iterate on their queries (Resilient Cyber).
  • OWASP ranked Excessive Agency third in its 2026 LLM list because autonomous systems often choose their own tools and expand paths beyond the original task (Rod Trent from Rod’s Blog).
  • The OCC’s GENIUS Act final rule, effective January 18, 2027, establishes new requirements for reserves, custody, and oversight of entities within its jurisdiction (insights4vc).
  • Brian Armstrong manages a system of always-on agents for Coinbase, using a Chief of Staff agent to oversee focus areas and present findings for human decision-making (Rich Holmes from Department of Product).

Try today · 45 min

Extract the tool-selection logic from your most critical agent into a deterministic code-based router.

Identify the top four user journeys your agent performs and hard-code the tool sequence for each.

You need

  • Current agent workflow documentation
  • List of tools the agent calls
  • Code editor

Steps

  1. Identify the top four user journeys the agent currently handles.
  2. Map the specific tools required for each journey.
  3. Write a conditional script in your agent's controller to trigger these tools in a fixed sequence.
  4. Remove the model's ability to choose tools for these specific journeys.
  5. Test the script to ensure the agent follows the hard-coded path.

One email instead of your newsletter pile.

Forward the newsletters you already get (a Gmail filter does it) and tell Hebel your goals. Every morning it reads them with 20,000 others and sends one short email, with one small thing to do today. Fewer emails, not more. Free during beta.

Start my paper →
LinkedIn X

Sources

  1. Resilient Cyber on hebel. →This analysis details indirect prompt injection vulnerabilities and Uber’s production defense-in-depth stack for managing agent intents.

  2. Rod Trent, Rod’s Blog on hebel. → ✎ Written by a personThis post explains why non-deterministic agent behavior and excessive agency complicate traditional security audits and path reconstruction.

  3. AAIF User Community on hebel. →These notes from AAIF Chennai Connect highlight the practical benefits of using deterministic code to control tool selection and sequencing.

  4. insights4vc on hebel. →This report covers the economic test for agent delegation and upcoming regulatory requirements for AI oversight in financial workflows.

  5. Rich Holmes, Department of Product on hebel. → ✎ Written by a personThis overview explains how leaders like Brian Armstrong structure personal agent systems to manage decision queues and reduce context switching.

More in Agentic Engineering

Build production agents or buy off-the-shelf tools

Production agents require evidence loops instead of green dashboards

Teams replace code review with automated intent verification

Anthropic and agents use 'dreaming' to curate memory

AI Agent Management yesterday

Autonomous agent coordination creates new management friction

Sign in to Hebel first

Likes, your collection and done Hebel need an account. We email you a six digit code; no password.