Yadda 3.0.0: BDD in the Age of AI Agents

Who Should Read

A developer running several coding agents who needs a way to verify the output actually matches the requirements

Core Mechanics

• Claude wrote most of Yadda 3.0
  • The modernisation used Claude Code with Opus 4.8 with very few human interventions.
  • It took roughly a day from start to npm publish while the author did other work in parallel.
• The work was split into phases
  • The phases were removing obsolete features, toolchain, mechanical formatting, source modernisation, API exploration, examples and CI, then docs and TypeScript definitions.
  • Each phase was planned before implementation.
• Code and tests were never changed in the same step
  • If an agent changes both, a green suite is weaker evidence because it can redefine what counts as correct.
  • The existing comprehensive test suite acted as a firm external constraint.
• The bottleneck is human coordination, not coding
  • The author could comfortably run three tasks at once, sometimes four or five.
  • Beyond that, context on each agent was lost, so he sees orchestration as the next key layer.

Community Reactions

• Opinion questioning why BDD uses natural language
  • The abstraction layer is good, but natural language seems an odd and expensive choice.
  • The commenter saw non-technical people use specs only once, with a traditional DSL.
• Experience adding BDD to AI codebases, mostly positive
  • It was added to two codebases, one engineered and one vibecoded.
  • High-level API and UI tests forced a first cleanup, such as standardised DB access and a single API SDK.
• Opinion that the spec-from-discussion flow is brilliant or a snake eating its tail
  • Turning project discussions into BDD rails for AI could be either, but is worth a try.

How to Apply

• Turn one meeting note into a candidate spec
  • Pick a requirements doc or meeting note today and have an LLM rewrite it as Given/When/Then scenarios.
  • Edit the terms yourself so they match your domain language.
• Separate code and test edits in agent tasks
  • When asking for a refactor, split production-code changes and test changes into separate steps.
  • Put a rule in the prompt that existing tests must not be edited during the code step.
• Break large jobs into phased epics
  • Order phases as removal, toolchain, formatting, source changes, docs, and ask for a plan before each.
  • Commit mechanical formatting separately from behavioural changes.
• Start parallel agents at three
  • Use git worktree to give each agent an isolated copy.
  • Note where you start losing context and set your parallel limit there.

Try It

# Paste into a Claude or ChatGPT chat (simplified version of the article's idea: requirements → executable spec draft)

Turn the requirements below into a BDD spec draft. Use domain terms consistently, and write sentences that describe intent rather than procedural scripts. Use GitHub-flavoured Markdown.

# Feature: feature name
## Scenario: scenario name
- list of sentences

Requirements:
{{meeting notes or requirements}}

At the end, list any inconsistent or ambiguous terms separately.

Next step: give the accepted spec to an implementation agent and tell it to write code that satisfies the spec without modifying tests.

Terminology

Related Papers

• Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy

  Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.

• Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents

  Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.

• Harness Engineering 101: The LLM is the brain, the harness is the body

  An agent is a loop managing a JSON array; 16 chapters build one from scratch

• Harness2 by Google Research – Recursive Agent Harnessing

  Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.

• Ousted Savannah police officer gave ICE agent husband access to Flock system

  A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired

• Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling

  Running Claude Code agents in tmux works well, but limits appear around four agents.

Related Resources

• http://www.stephen-cresswell.com/2026/08/15/Yadda-3.0.0-BDD-in-the-Age-of-AI-Agents.html