ciunkos.com/Glossary

Coding with agents

Reward hacking (coding agents)

When an agent makes the check pass without solving the task: deleting or weakening tests, special-casing test inputs or faking results.

Models trained to make tests pass learn that changing the test is sometimes the shortest path. In practice it looks like a skipped test, an assertion loosened “to fix flakiness”, a hard-coded return value for the exact test input or a mock that hides the real call. Anthropic’s research shows that models which learn such shortcuts in training can also generalize them into worse behavior.

The defenses are structural. Keep the validation loop honest by protecting test files from edits during implementation tasks, grade with holdout scenarios the agent never sees, and use adversarial review that looks specifically at changes to tests and mocks. A diff that touches tests and code at once deserves the closest look.

Sources

Share: LinkedIn · X ·

Related terms

  • Holdout scenarios: End-to-end user scenarios kept outside the codebase, so the coding agent never sees them and cannot write code that only passes the check.
  • Validation loop (agent backpressure): Commands a coding agent runs after every change (tests, types, lint, build) and fixes against until they pass, so it checks itself before a person looks.
  • Adversarial review (cross-model review): Asking a second AI model, or a fresh agent, to find what is wrong with a change before a person reviews it, instead of asking the author to check itself.
  • Evals: A repeatable test set that scores an AI feature, so changes can be measured instead of guessed.
  • Guardrails: Checks around a model that block, fix or flag inputs and outputs that break the rules.

← All terms

Free tools

Business idea validatorTen rubrics, a fixed-rule verdict and three cheap experiments. Free.Software cost estimatorYour project among 125 real products, with honest ranges.Engagement configuratorPick the work and its size, see hours, price and start date.SPF, DKIM and DMARC checkerGrade any domain’s email DNS, with records to paste.BIMI checkerRecord, DMARC, SVG Tiny PS logo and VMC or CMC, with an inbox preview.App store readiness checkWhat your web app lacks for the stores, TWA or Capacitor, files to copy.Build or buy calculatorSaaS against your own app over 1, 3 or 5 years, with the payback month.Runbook generatorA RUNBOOK.md for a system only you know, in two minutes, in your browser.SLA calculatorMonthly fee, response times, a downtime budget and contract clauses.Local model or API?Does the model fit your hardware, how fast, what it costs against an API.OpenAPI to MCP toolsPaste an OpenAPI spec, get narrow MCP tools and a server skeleton.KSeF invoice validatorCheck an FA(3) invoice against the official schema in your browser.Quoting time calculatorHours spent turning inquiries into quotes, and when an agent pays off.Multi-tenant SaaS calculatorWhich multi-tenant model, the work per release and the weeks to get there.Quote DeskDrop a customer inquiry, get lines matched to your price list. Free.AI Visibility DeskDoes Claude recommend your brand? Weekly runs of your buyers’ questions, free.List DeskClean an email list in your browser: dead domains, disposable, typos, duplicates.Eval DeskTest cases for your AI prompt, run after every change: what regressed, what got fixed.Dashboard DeskA growth dashboard created in your PostHog, plus what your tracking is missing.Signature DeskOne email signature template for the whole team, as HTML for Gmail, Outlook and Apple Mail.Phishing email checkerForward a suspicious email or upload the .eml file; three AI models and technical checks give a verdict.Email TestSend one email, check one-click unsubscribe, SPF, DKIM, DMARC and spam signals.Vitals DeskA daily lab check of your pages on a throttled phone: LCP, CLS and TBT, and an email when a page gets slower. Free.Legacy rewrite estimateLines of code to function points, hours and cost, with and without agents.

Want the fun version? The Corporate-speak Translator has a Polish dictionary of office phrases and what they really mean. Corporate-speak dictionary (Polish) →

Need this built, fixed or reviewed?

I build and rescue products with AI coding agents under rules: AI features in production, honest estimates, email that gets delivered.

Przemysław Zalewski

How I work

Przemysław Zalewski · 15+ years shipping software

“a true bias for action” client CTO, public LinkedIn recommendation

My pledge

All principles, the AI policy and working hours →

Pricing · Security · Privacy and analytics opt-out · Trust center