Skip to content
Laith Al-Saadoon
Writing

Frontier Field Notes

Writing from the autonomous work horizon: how long an agent can work on its own, what it needs to get there, and how to tell when it's slipping. Field Notes are the long-form essays, numbered as a series. Side Quests are the short detours that didn't need one.

Long form

Field Notes

9 essays

  1. Prove the Spec Before You Write the Code with symspec

    Find contradictions in structured requirements before writing code. symspec uses Z3 to prove conflicts and reports the limits of what it checked.

    July 9, 20269 min readSpec VerificationCoding AgentsSMTEARS

  2. Yield Engineering for Fleets of Autonomous Coding Agents

    Shipping more pull requests is not the same as delivering useful work. Apply yield, rework, and defect measurements to a fleet of coding agents.

    June 11, 20267 min readYieldMetricsAutonomous Delivery

  3. Self-Improving AI Agents Will Reduce the Code You Ship

    Given one sentence, an agent pulled data, ran statistics, wrote a brief and kept what it learned in six minutes. Most software we ship is plumbing it can write.

    June 4, 20266 min readSelf-Improving AgentsToolsArchitecture

  4. ERPAVal, a Gated Six-Phase Workflow for Claude Code Agents

    ERPAVal runs Claude Code agents through six gated phases and writes what each session learned to disk, so the next run starts with yesterday's lessons.

    May 14, 202615 min readAutonomous DevWorkflowClaude Code

  5. Agent Evals That Show When Your AI Agent Is Getting Worse

    Catch regressions before deployment and bad output during a run. A comparison of seven eval frameworks and the feedback loops that make them useful.

    March 19, 20266 min readEvalsRegressionQuality

  6. Why Coding Agents Need a Computer That Isn't Your Laptop

    The coding agent harness is absorbing knowledge work, and laptops fail it at team scale. Agents need isolated, durable, scalable cloud compute to run for hours.

    March 12, 20267 min readAgentsInfrastructureCloud

  7. The Autonomous Work Horizon Is Doubling Every Five Months

    A March 2026 analysis of METR task horizons: a 4.8-month doubling trend, its reliability limits, and why human task duration is not agent runtime.

    March 5, 20266 min readAutonomyBenchmarksMETR

  8. Closing the Loop on AI-Generated Code with Automated Review

    Independent AI review can catch design problems that linters miss. Build a review loop that complements tests, static analysis, and human judgment.

    February 12, 20266 min readCode ReviewQualityAI-Generated Code

  9. Why Frontier AI Labs Are Obsessed with Coding Abilities

    Code is the most general interface to the modern economy. It reaches APIs, terminals, files, and databases, so labs build agents that turn intent into action.

    January 8, 20265 min readFrontier ModelsCoding AgentsEconomics
Short

Side Quests

3 detours