Talk ↗Tooling, Workflows, Process
A field perspective on AI coding tooling, workflows and processes, drawn from terabytes of real coding-agent session logs. Given at the Rome2rio dev day.
Resources
Practical notes on AI coding
Configuration, cost, and workflow — written as the tools change, and dated so you know what era they belong to.
Token counts are about to become the new lines of code
We already learned that measuring commits and lines of code doesn't measure effectiveness. Token-spend dashboards are about to teach us the same lesson again, with a dollar sign attached.
costengineering-productivitymetrics
Retuning your setup for GPT-5.6: seven changes that save tokens, dollars and time
A new major model release means your old configuration is probably wrong. What we learned from the GPT-5.6 system card and early field reports - and the specific settings to change.
gpt-5.6codexconfigurationcost
The Headroom teardown: zero tokens saved, then 88% - depending entirely on how you wire it up
A token compressor claimed 12% savings while delivering nothing to the actual session. Wired up properly, the same tool does something genuinely clever. The difference is entirely in one architectural detail almost nobody gets right.
context-managementcosttoolingtokens
Big diffs are fine, actually
The size of a diff isn't the problem - the distance between lines of code and what actually happened is. Review the decisions and the plan, not the line count, and diff size stops mattering.
code-reviewagentic-codingworkflow
Talk ↗Agents Lie. The Compiler Doesn't.
How Chris builds Rust he'd actually ship — with a small fleet of coding agents, and a compiler that keeps them honest.
The loop that cost $1000+ overnight - and why 'improve test coverage' was the actual bug
An open-ended goal, no stop condition, and a cron job left running overnight turned into 30+ PRs and a four-figure token bill. What designing a loop properly actually requires.
agentic-codingcostautomationguardrails
Talk ↗12TB of AI coding agent logs — what works, what fails
Empirical findings from Cadence session-log data, given at AI Engineer (Melbourne).
CI isn't a gate anymore - it's a verify tool the agent calls
Continuous Integration used to mean one build per branch, run once you were done. In agentic workflows it's part of the core loop, called repeatedly, and that changes what your test suite needs to look like.
citestingagentic-codinginfrastructure
Four levels of working with a coding agent (and where the value actually is)
Vibe-coding, the 'you're so smart, do everything' challenge, tiny unjoined-up tickets, and building-without-learning - a rough taxonomy for finding where you actually sit.
agentic-codingengineering-productivityworkflow
Talk ↗Notes From the Field
Dave Slutzkin (CEO, Cadence) on what's actually happening inside AI coding workflows.
RIP token maxxing: the subsidy on programmatic Claude usage is ending
Starting June 15, every GitHub Action, cron job, and SDK-built agent quietly running on your Pro plan draws from its own metered credit. The arbitrage that funded a generation of agentic workflows is over.
costclaudeautomationgovernance
The tanstack worm wasn't an attack on agentic coding. It was an attack built for it.
It hid in .claude/ hooks to survive reboots and signed its commits as the Claude Code bot. The real finding isn't whether you dodged this specific worm - it's whether you'd know if you hadn't.
securitysupply-chainagentic-codinggovernance
A planning workflow that works: spec, design doc, implementation plan
Best practice right now is spending 50%+ of your time planning before an agent writes a line of code. Here's the three-document structure we see working across teams.
planningagentic-codingworkflow
Line-by-line code review is finally dead. Here's what replaces it.
AI coding tools broke code review by exploding PR volume. The fix isn't asking developers to review more - it's being honest about which parts of review a human still needs to do.
code-reviewagentic-codingworkflow
Talk ↗Dev Productivity in the Glittering AI Future
What we've seen across terabytes of coding agent logs, and three things that guide effectiveness.
Beyond AGENTS.md: Harness Engineering, Loop-Based Delivery, and Context-Aware Prompting
Why static agent instruction files aren't enough for production AI coding - and how harness engineering, loop-based workflows, and context-aware prompting create systems that actually ship.
ai-codingengineering-productivityagents
Tunnel vision: why agents get stuck, and how to break them out
Agents lock onto a course of action and stop questioning whether it was right in the first place. The smells to watch for, and why it happens more the longer a session runs.
agentic-codingworkflowdebugging
"Backpressure" is testing, rebranded - and it matters more than ever
QA quietly disappeared over the last two decades of agile and DevOps. AI coding is bringing it back under a new name, because agents are far worse than humans at knowing when they're wrong.
testingagentic-codingci
For AI coding to work well, three things have to line up
The model, the engineer driving it, and the codebase underneath it. Get one wrong and the other two can't save you - notes from reviewing a few terabytes of real coding sessions.
agentic-codingengineering-productivitycodebase-health
The subagent orchestration prompt I keep reusing
A concrete implement/review/triage loop for Claude Code subagents that can run unattended for hours on a TODO.md - the actual prompt, and why each piece of it exists.
subagentsclaude-codeagentic-codingprompts