Blog

Notes from the workshop.

How we build agents that ship real software: engineering write-ups, product decisions, and what we're learning about codebases in the wild.

Code-grounding

Why agents need to read your codebase before they write to it, and what reliable looks like in practice.

How to make an AI coding agent reliable on a real codebase

Reliable AI coding agents come from grounding, evaluation, verification, and review, not a smarter model. A practical guide to trusting agents on real code.

JUL 14, 2026ENGINEERINGNot everything that fails is a failure: blocked vs wrong JUL 12, 2026ENGINEERINGGetting your codebase agent-ready JUL 10, 2026ENGINEERINGContext engineering for coding agents JUL 8, 2026ENGINEERINGDocumentation drift is a bug, treat it like one JUL 6, 2026ENGINEERINGBuilding an eval harness for a coding agent JUL 4, 2026ENGINEERINGGrounding beats prompting: give a coding agent the right context JUL 2, 2026ENGINEERINGCatching hallucinated file paths cheaply JUN 30, 2026ENGINEERINGImpact analysis before you touch a line JUN 28, 2026ENGINEERINGLiving documentation: docs that keep up with the code JUN 26, 2026ENGINEERINGLLM-as-a-judge: when to trust a model to grade a model JUN 22, 2026ENGINEERINGRAG for code: when retrieval helps an agent and when it hurts JUN 20, 2026ENGINEERINGWhy an agent should read your tests before your code JUN 18, 2026ENGINEERINGStructured output guarantees the shape, not the truth JUN 16, 2026ENGINEERINGAcceptance criteria an agent can actually test against

Trust and oversight

Approvals, audit trails, and isolation: what has to be true before an agent is allowed near production.

Can you trust an AI agent to ship to production?

Trustworthy AI agents earn production through oversight, accountability, least privilege, and isolation. A guide to what to demand before one ships your code.

JUN 14, 2026PRODUCTWhy an agent should never approve its own work JUN 12, 2026PRODUCTRunning an AI agent on code you do not own JUN 10, 2026PRODUCTThe audit trail is the product, not a feature JUN 8, 2026PRODUCTAutonomous vs unsupervised: the distinction that matters JUN 6, 2026PRODUCTConfidence-based oversight: match the gate to the risk JUN 4, 2026PRODUCTData isolation for AI tools that touch customer code JUN 2, 2026PRODUCTData residency for AI tools: bring your own cloud MAY 31, 2026PRODUCTLeast privilege for AI agents MAY 27, 2026PRODUCTPrompt injection when your agent reads untrusted content MAY 25, 2026PRODUCTReconstructing an AI decision after an incident MAY 23, 2026PRODUCTSecrets management when an agent touches your systems MAY 21, 2026PRODUCTSingle-tenant vs multi-tenant when isolation is the promise MAY 19, 2026PRODUCTWhat SOC 2 asks about your AI agents MAY 17, 2026PRODUCTWhat \"a human approved this\" should actually mean MAY 15, 2026PRODUCTWhich decisions must stay human when an agent ships code MAY 13, 2026PRODUCTWho is responsible when an AI ships a bug?

Architecture

How to shape an agent system: handoffs, state, concurrency, and the parts that should stay plain code.

AI agent architecture: designing a system that survives production

AI agent architecture is what separates a demo from a system you can run. A guide to designing agents for reliability, model fit, and scale, without the hype.

JUL 16, 2026ENGINEERINGDesigning agent handoffs and contracts JUL 14, 2026ENGINEERINGConcurrency and rate limits at scale JUL 12, 2026ENGINEERINGThe context window is an architecture constraint JUL 10, 2026ENGINEERINGDesigning for model swaps: staying provider-agnostic JUL 8, 2026ENGINEERINGWhen a step should be deterministic code, not an agent JUL 6, 2026ENGINEERINGIdempotency and retries when your worker is an LLM JUL 4, 2026ENGINEERINGOne product, many repos: planning across service boundaries JUL 2, 2026ENGINEERINGOrchestration vs autonomy: who is in control JUN 28, 2026ENGINEERINGPreventing duplicate work when a task fires twice JUN 26, 2026ENGINEERINGResumability: surviving an interruption mid-task JUN 24, 2026ENGINEERINGSmall specialized agents beat one autonomous mega-agent JUN 22, 2026ENGINEERINGState machines for agent workflows JUN 20, 2026ENGINEERINGStreaming structured output: parsing an agent while it thinks

Economics

What agentic coding actually costs, where the money goes, and how to keep the bill predictable.

The economics of AI coding agents: what they cost and how to control it

AI coding agent cost is not a mystery. A grounded guide to what agents cost per task, the levers that cut the bill, and the business model underneath it all.

JUL 15, 2026COMPANYAgent loops and quadratic token growth JUL 13, 2026COMPANYAI product margins: why COGS is back JUL 11, 2026COMPANYBatch the boring half of your pipeline JUL 9, 2026COMPANYBring your own model: pricing without token-margin risk JUL 7, 2026COMPANYBuild vs buy an internal agent platform JUL 5, 2026COMPANYCapping runaway agent spend JUL 3, 2026COMPANYHow to price an AI product JUL 1, 2026COMPANYIs an AI coding agent worth it? A cost-benefit read JUN 29, 2026COMPANYModel routing: send the cheap work to the cheap model JUN 25, 2026COMPANYPrompt caching: the cheat code for repo-grounded agents JUN 23, 2026COMPANYThe token math of an agentic coding task JUN 21, 2026COMPANYWhat an AI coding agent actually costs per feature JUN 19, 2026COMPANYWhy output tokens dominate your bill

Problems

The failure modes teams hit with AI coding tools, named plainly and traced to their cause.

Why AI coding agents fail in production (and what actually fixes it)

AI coding agents fail in production for predictable reasons: weak grounding, no verification, no accountability, brownfield reality. A map of the failures.

AUG 4, 2026KNOWLEDGE BASEGreenfield bias: why AI coding tools assume a clean slate you do not have AUG 3, 2026KNOWLEDGE BASECost surprise: why AI coding bills are unpredictable and how to see them coming AUG 2, 2026KNOWLEDGE BASEThe rework tax: the hidden cost of AI code that looks done but is not AUG 1, 2026KNOWLEDGE BASEThe trust gap: why security-conscious teams will not run agents on their code JUL 31, 2026KNOWLEDGE BASEWho approved this? The accountability gap in AI-assisted shipping JUL 30, 2026KNOWLEDGE BASEThe review bottleneck: when AI writes faster than humans can check JUL 29, 2026KNOWLEDGE BASECross-boundary breakage: changes that pass locally and break elsewhere JUL 28, 2026KNOWLEDGE BASEDependency blindness: agents that ship blocked or out-of-order work JUL 27, 2026KNOWLEDGE BASEThe last mile: a pull request is not a shipped feature JUL 26, 2026KNOWLEDGE BASEContext rot: why agents lose the plot on large or legacy codebases JUL 25, 2026KNOWLEDGE BASEPlausible but wrong: the most expensive failure mode of AI-written code JUL 24, 2026KNOWLEDGE BASEThe demo-to-production cliff: why the impressive demo does not survive real work

Earlier posts

Notes from before the current pillars took shape.

JUL 15, 2026KNOWLEDGE BASEWhy our agents read your code before they read your PRD