DEV PRODUCTIVITY
in the glittering AI future
what we've seen in ? TB of coding agent logs.
Dave Slutzkin - CEO Cadence
2026: Everyone's using AI coding tools.
Three things guide effectiveness and productivity:
Models/Tools
Way too soon to standardise, unfortunately.
Definitely don't use Copilot.
Developer
How you work matters but everyone's working differently. No best practice yet.
Codebase
Legacy code is fine, but docs and tests make a huge difference to the agents.
State-of-the-art this week
codex + gpt-5.4
% of AI coding sessions with successful outcome
After
vibe coding
How do we make this engineering?
Three modes of AI coding
Autocomplete
10–20% faster?
Copilot/Cursor hegemony. Broken tab keys.
API compatible with previous workflows.
Vibe coding
10x faster?
Claude Code is the new emperor. Fast, sloppy, liberating. Probably unmaintainable.
Agentic engineering
3x faster?
Autonomous agents guided by real specs, tests, and guardrails. Production quality.
Mind-meld with the agent(s).
- Long planning phases: hours working with the agent asking and answering questions
- Tests first: agent eagerly writes tests, and you review them while it builds
- Most testing on CI: local testing is targeted; CI does e2e and integration
- Small tasks in parallel with a big task: bouncing between tmux panes
- Lots of automated PR review: and the agent is instructed to read and triage review points
PRs per week
Average hours to merge PR
Backpressure and docs
Help the agent with the codebase
Where we need to get our legacy code to
- Heavy integration and e2e testing: but they run on CI (optimised, parallel) probably not locally
- Extensive documentation: AGENTS.md hierarchy with progressive disclosure; deep and up-to-date architecture docs; mermaid diagrams
- Harness engineering: watch the agent for dead ends and failure then give it the tools it expects; don't burn tokens on success output
Three things to take back to your team
Models/Tools
It changes every few weeks. Right now: codex + gpt-5.4. Don't standardise yet.
Developer
Mind-meld with the agent. Long planning, tests first, small parallel tasks, automated PR review.
Codebase
CI-heavy testing, deep docs with progressive disclosure, harness engineering to unblock agents.
teamcadence.ai
