Free guide · 5 failure modes

Why context engineering isn't enough

5 failure modes of coding agents — and how to build the harness that stops them.

PDF · no email required

Download the guide. Slide 2: Your agents got better. Trust didn't. You've engineered the context; your process hasn't kept pace (96% of engineering leaders say theirs can't keep up with their agents); harness engineering restores trust.

What's inside

Five failure modes, each with its fix.

AGENTS.md and skills steer the agent, but they're probabilistic. The harness is the deterministic layer around it: shared workflows, architecture rules, sandboxes and CI checks that decide what ships.

  1. Slide 3, failure mode 1: green in CI, broken in production. 82% of tech leaders had a production failure tied to AI-generated code.
    1 · Green CI, broken prod

    Green in CI, broken in production.

    The harness: let deterministic checks decide what ships.

  2. Slide 6, failure mode 2: nobody understands the system anymore. 45% of developers would need AI's help to explain code they wrote.
    2 · Comprehension debt

    Nobody understands the system anymore.

    The harness: get flagged when your picture of the system goes stale.

  3. Slide 9, failure mode 3: agents delete what nobody asked them to. Five incidents, from Replit to PocketOS.
    3 · Destructive agents

    Agents delete what nobody asked them to.

    The harness: wrap the LLM in deterministic guardrails: move rules from context into code.

  4. Slide 12, failure mode 4: what your agent can reach, an attacker can too. The lethal trifecta: untrusted content, private data and a way out.
    4 · Hijacked agents

    What your agent can reach, an attacker can too.

    The harness: limit what gets into the agent, what it can reach and what can leave.

  5. Slide 14, failure mode 5: review is drowning your best engineers. 98% more pull requests merged, 91% more time spent in review.
    5 · Review overload

    Review is drowning your best engineers.

    The harness: build less, and automate review one step at a time.

The PDF ends with three fixes you can make this week, each with the exact config, and a 5-minute checklist: tick what your team can demonstrate today, then start with the gap that could hurt most.

Free 20-minute call · with Renato

Your first fix, in writing.

Tell me how your team ships agent-written code. Within two business days you get one written page.

Not for you if your team is still choosing a coding agent or running a pilot: this is for teams already shipping with one.

Book your free call I work with 3 teams at a time. 3 spots open.

Prefer to write? Send me a message.

  • Your top riskThe gap most likely to hurt you, judged against the five failure modes.
  • Your first fixA concrete step your team can take this week.
  • Do you need help?Which paid option fits, or why you don't need one.