Kiln - An Operating System for Managing Fleet Of Agents In an AI-Native Organization

Kiln-Code is an operating system that covers the full lifecycle of managing AI coding agents in production, from messy requirements to shipped software, with compound learning across iterations.

I built it because managing a fleet of agents (Claude Code, Cursor, Codex) across multiple products exposed a recurring problem: the bottleneck is never the agent. The bottleneck is everything around the agent. How you translate requirements into something an agent can execute correctly. How you validate technical feasibility before committing to a full build. How you catch the specific words in your prompts that reliably cause agent misbehavior. How you review agent output when you are the only reviewer. How you diagnose failures when an agent produces wrong results. How you prevent your specifications and your code from drifting apart over time.

Kiln-Code systematizes all of this into a gstack like library that runs inside the tools engineers already use.

What I will show live:

1. Intake - paste a raw, unstructured requirement (voice message, brain dump). System extracts structured intent: core problem, entities, risks, scope boundaries.
2. Spike - before committing to a full build, run a time-boxed R&D exploration. The system generates a constrained experiment spec: what hypothesis to test, what the agent is allowed to explore, what it must not touch, how long it gets, and what the structured findings report must contain. The agent explores within these boundaries and produces a feasibility verdict with evidence. I will show how a spike prevents you from committing engineering weeks to an approach that a two-hour controlled experiment could have killed early.

3. Spec generation - system produces a structured specification with nine sections covering intent, domain model, architecture, operations, engineering norms, hard constraints, and codebase context. Spike findings feed directly into the spec if you proceed to build. I will show how different task types (bug fix vs new feature vs refactor) activate different sections.

3. Hardening - the linguistic optimization pipeline runs on the spec. I will show it catching trigger words that cause known agent failures, flagging naming inconsistencies, and detecting scope bleed where operations reference code outside the defined boundaries.

4. Compilation - same spec compiles into structurally different prompts for Claude Code versus Cursor. I will show the two outputs side by side and explain why each difference exists (compaction behavior, IDE context, failure tripwire placement).

5. Agent execution - dispatch the compiled prompt to Claude Code live and watch it execute.

6. Failure diagnosis - I will deliberately trigger a known agent failure pattern, then classify it against the documented taxonomy and walk the fix path live.

7. Sync - show the spec updating after code changes, and code regenerating after spec changes. Bidirectional. The spec is always the source of truth