Currents
Where enterprise software is heading, from the people building it.
GitHub Copilot Defaults Flip Oct 22: 5 Settings to Make Now
GitHub Copilot turns on unconfigured features October 22, 2026. Five settings to make first: the global default, MCP servers, models, and org admins.
Read post →
Build vs. Buy a Coding Agent: Four Questions to Ask First
Build vs. buy an AI coding agent: Ramp built its own, but most mid-market teams should buy and own a thin layer on top. Four questions to settle it first.
When Should an AI Agent Stop and Ask a Human? Four Gates
An AI agent should stop for human approval before it publishes, does anything irreversible, moves money, or edits its own rules. Why those four, and only four.
5 Things AWS Key Quarantine Won't Do When Your Key Leaks
AWS key quarantine doesn't disable a leaked access key or block role pivots. Five gaps the AWSCompromisedKeyQuarantineV3 policy leaves, and what's still on you.
The DuckDB Team Joined AWS. What Actually Changes for You.
DuckDB stays open source under the MIT license after its team joined AWS, and a nonprofit foundation still holds the IP. What that locks in and what to watch.
AI Guardrails Don't Need to Be Hacked, Just Asked Nicely
Cisco Talos read real attacker prompt logs: most AI jailbreaks were plain claims of permission. Build guardrails that verify authorization outside the model.
Is Code Review Going Away? What Replaces It for Small Teams
Line-by-line code review is fading as agents write more code, but verification is not. The four checks that replace reading every diff, and what they cost.
What Zero Data Retention Actually Buys You
Zero data retention is several promises under one name. What OpenAI's fine print covers, where a 30-day record still lives, and five questions for AI vendors.
5 Signs Your AI Project Has a Data Problem, Not a Model One
Five signs your AI project has a data problem, not a model problem: a demo on hand-picked data, no data owner, over-broad access, and sources that disagree.
IT Skills AI Is Devaluing, and the One It Hasn't
ADP and Stanford payroll data show which IT skills AI is devaluing: routine execution. Judgment, design, and directing technical work are tied to higher pay.
Coding Agent Lock-In: 5 Portability Checks Before You Commit
How to avoid coding agent lock-in: five portability checks on instructions, repo knowledge, integrations, data export, and exit cost before you commit.
AI Feature Adoption: Why Customers Ignore What You Shipped
51% of software operators say fewer than one in four customers use their AI features. Why shipped AI goes unused, and the workflow-first fix for adoption.
5 Things to Check Before AWS Ends Email Certificate Validation
AWS stops renewing email-validated ACM certificates on September 30, 2027. Five checks to run now so a quiet renewal failure doesn't take a site down.
How to Reduce AI Token Costs: Fix the Architecture First
To reduce AI token costs, fix the pipeline before the invoice. Spend caps ration the bill; model routing, context, caching, and code structure set it.
Ruby Marshal RCE: 5 Things We Check in a Legacy Rails App
A universal Marshal RCE chain hits Ruby 4.0.6 and back to 3.3. Five places Marshal deserialization hides in a legacy Rails app, and what we check first.
Rails AI Agent Benchmark: What the Scores Can't Predict
The first Rails AI agent benchmark reads like a leaderboard. The useful signal is API recall, and what an atomic score can't predict about an older Rails app.
TDD With Coding Agents on Legacy Code That Has No Tests
Should coding agents use TDD on legacy code with no tests? Not first. Pin current behavior with characterization tests, then let the agent test-drive changes.
AI Coding ROI: The Maintenance Cost Nobody Prices In
AI coding ROI models count speed and review time, not who will understand the code later. On older systems that maintenance cost compounds. How to price it.
Refactoring ROI Just Got a Price Tag, Measured in Tokens
Coding agents bill by the token to read your code, so refactoring ROI is now measurable. One experiment cut input tokens 83 percent. How to price yours.
5 Ways Sloppy Code Inflates Your AI Coding Bill
Yes, messy code raises your AI coding bill. Agents pay to read every line, and one refactor cut input tokens 83 percent. Five things we check first.
Coding Agent Security: 5 Checks Before It Touches a Repo
Wiz's GhostApproval research tricked six coding assistants with one symlink. Five coding agent security checks we run before an agent touches any repo.
Building a Mac App With Claude: What We Did and What It Did
Claude wrote nearly all of Tinystat, our free Mac menu bar app. Architecture, review, and testing stayed ours. What that split means for scoping AI coding work.
AI Usage Policy: 5 Things It Should Actually Cover
What an AI usage policy should cover: banned data, what a share link exposes, vendor retention settings, who approves new tools, and what happens on a slip.
The Forward-Deployed Engineer Gap in the Mid-Market
Forward-deployed engineers embed with your team to ship AI to production. AWS and Microsoft sell the model to big accounts. What mid-market buyers should do.
AWS Account Lockout: The Migration Risk Nobody Rehearses
Migration plans rehearse the cutover and skip account access. How to keep a suspended or orphaned AWS account from becoming an outage you cannot end.
How to Contain AI Agents Without a Platform Team
Approval prompts will not contain an AI agent. What a small team can do instead: a scoped identity, staging access, egress rules, and a throwaway workspace.
Cutting IT Maintenance to Fund AI Caps What the AI Returns
Many CIOs now fund AI by cutting maintenance and legacy contracts. Deferred maintenance is a loan, and the neglected systems cap what the AI can return.
AI Coding Moved the Bottleneck to Code Review. Staff for It.
AI coding tools raise code volume, but delivery stalls when code review becomes the constraint. How to staff and sequence a team for review capacity.
How to Scope AI Work When Leadership Expects 100x
When leadership already promised a 10x or 100x AI productivity gain, measure accepted work, not generated code, and scope against the stage where work waits.
AI Agent Credentials: What to Audit Before Production Access
Most AI agent security failures are old plumbing mistakes: broad credentials, unvalidated URLs, secrets in logs. The audit to run before production access.
Legacy Systems and AI Agents: Fix the Foundation First
An AI agent amplifies the legacy system under it. The four fixes to make first, from data model to dead code, and a checklist to tell when you are ready.
Modernizing ColdFusion With AI: Evidence Beats Plausibility
AI speeds up legacy ColdFusion modernization only when logs, config exports, and characterization tests check every claim it makes about your code first.
Why Long-Running AI Agents Fail and What the Harness Must Do
Long-running AI agents fail because they lose state between context windows, not because the model is weak. What a working harness does and what to ask.
Evaluating New AI Model Releases Like a Buyer, Not a Fan
How to decide if a new AI model release matters to your business: check three deltas, test on your own eval set, and review quarterly instead of on launch day.
Why AI Agents Stall in Production. It Is Not the Model.
Most AI agent projects never scale to real value. The cause is rarely the model. It is the missing tracing, evaluation, guardrail, and monitoring layer.
ColdFusion 2025 AI Framework: Should Your Legacy App Use It?
ColdFusion 2025 Update 8 added native LLM, RAG, and MCP support. Whether your app should use it depends on one question: is it staying on ColdFusion?
Medical Practice Revenue Down? How We Found $123K in Aged AR
When practice revenue falls but patient volume holds, the cause often hides in aged AR. How AI-assisted analysis found $123K in hours, and what it did not do.
How to Run an AI Pilot That Doesn't Waste the Budget
Most AI pilots end in a demo, not a decision. How to run an AI pilot that ends in one: aim it at your binding constraint, set a baseline and kill criteria.
Agentic Arbitrage: What Gartner's $234B Forecast Assumes
Agentic arbitrage is AI agents doing work across your systems. Gartner puts $234B of SaaS spend at stake. Capturing it takes integration and retained context.
Claude Opus 5's Launch Weekend: Evaluate Now, Migrate Later
Claude Opus 5 shipped July 24, 2026, and launch weekend brought elevated errors. Evaluate it now, log the served model, and wait before moving production.
Moving a 200GB Legacy Codebase Into Git Starts With the Data
A legacy codebase full of reports, uploads, and backups won't fit in Git. Move the data out first, then version the code. Here is the order that works.
AWS vs Azure for Legacy Windows and ColdFusion Apps
AWS or Azure for legacy Windows, SQL Server, and ColdFusion apps? Both run them well. Who operates the cloud and your SQL Server licensing decide it.
Questions to Ask an AI Consultant Before You Sign
Seven questions to ask an AI consultant before you sign, to separate firms that ship production AI systems from firms that sell decks and workshops.
Taking an Inherited AI Proof of Concept to Production
An AI proof of concept that works is not mostly done. What gets rewritten on the way to production, how to contract for it, and the review to run first.
AI Coding Agents in Production: What Works and What Drifts
AI coding agents ship real production code when verification comes first, tasks stay small, and a human owns the merge. What holds up and what drifts.
Replacing Cron Batch Jobs With Rails and Temporal
Cron batch jobs fail silently and restart badly. How Rails plus Temporal durable workflows fix that, and how to migrate a legacy job without breaking month-end.
ColdFusion 2025 Migration Checklist: Notes From Production
What a ColdFusion 2025 migration needs beyond the code: a dependency inventory, a reporting plan, testing before the DNS change, and a rollback you can use.
ColdFusion in 2026: Still in Production, and the Real Risk
Yes, ColdFusion still runs insurance, government, and education systems in 2026. The real risk is not the language. It is unpatched, unmanaged servers.
API Keys Leaked on GitHub? What to Do in the First 24 Hours
Making the repo private does not fix a leaked API key. What to do in the first 24 hours after keys hit a public GitHub repo, in the order that matters.
ColdFusion Migration Strategy: Triage Before You Rewrite
A ColdFusion migration plan starts with triage, not a target language: sort the app into keep, rewrite, and retire, then use AI to map it, not to translate it.
Is Your Legacy App Worth Modernizing? A Decision Framework
Leave it running, remediate, modernize, or rewrite? Five questions decide what a legacy application deserves, before anyone proposes a fix or writes code.
What an AI Readiness Assessment Actually Tells You
An AI readiness assessment scores data, skills, and your tech stack separately to find the one constraint that will stop an AI project first. How to read yours.
Have a problem worth solving?
Tell us what you are trying to build or modernize, and we will tell you honestly how we would approach it.