The Digital Foundry
Find my order
GuideTDF-GD-02
Agents in Production
The failure modes nobody puts in the demo.
THE DIGITAL FOUNDRY$19

Guide · 02

Agents in Production

A field guide to running agents against real work instead of a demo: tool-call loops that don't terminate, context that silently rots, permissions that are either too loose or too strict to ship. Each failure mode gets a recognizable signature, a cause, a fix, and the guardrail that stops it recurring.

$19one-time · yours forever
Instant delivery — download page opens right after checkout

All sales final — files arrive instantly, so there’s nothing to return. Broken or not as described? We fix it: caremindcontactus@gmail.com

What’s inside

Open the folder before you buy it.

This is the exact manifest of the download — no mystery box. A full excerpt is free below, so you can judge the quality before paying.

Read the free sample →
agents-in-production/ · 13 files · 0.75 MB
START-HERE.mdplain text
agents-in-production.pdf44 pages
agents-in-production.epubebook edition
agents-in-production.mdplain text
failure-modes/5 files
infinite-tool-loops.md·
context-rot.md·
permission-scoping.md·
silent-partial-failures.md·
runaway-cost.md·
incident-postmortem-template.docxplain text
guardrail-checklist.mdplain text
guardrail-checklist.pdf3 pages

Free sample · Infinite tool loops

A composite incident: 40 retries in nine minutes.

A browser-automation step hit a page that returned a 200 status but rendered a CAPTCHA instead of the expected form. The parsing step found no recognizable fields, classified that as "form not found — retry navigation" instead of "unexpected content — escalate," and reloaded the same URL 40 times in nine minutes. What actually stopped it wasn't an internal safeguard — it was a rate limiter on the target site blocking the account's IP.

Post-incident, "no recognizable form fields" was reclassified from a transient to a structural failure category with a ceiling of one attempt, and CAPTCHA-shaped content was added as its own case that escalates straight to a human-review queue instead of retrying at all.

— excerpt ends. Four more failure modes, each with a cause, a fix, and a guardrail, in the complete guide.

Buy it if. Skip it if.

Built for you if

  • You have an agent past the demo stage and it's touching real user data or spend
  • You want named failure modes and fixes, not general agent theory
  • You're about to scope permissions for an agent and want the tradeoffs laid out first

Skip it if

  • You haven't built an agent yet — this assumes something is already running
  • You're only using single-turn LLM calls with no tool use — most of this won't apply
  • You want a specific framework's API docs — this is failure patterns, framework-agnostic

Questions