Skip to content

Articles

Practitioner-focused analysis bridging research findings to enterprise adoption strategies.


Measure the Work, Not the Meter

June 2026

Token billing itemized the cost of AI work and measures none of its value. Auditing token spend prices the new way of working with the old way of measuring. The answerable question is cost and cycle time per delivered change, against a baseline, and the baseline-capture window is open now.

Topics: Delivery economics, cycle time, cost per delivered change, baselines, gateway attribution, Goodhart's law

Read →


Harness vs Framework: Why 'Just Use LangChain' Stopped Being the Answer

June 2026

An agent is a model plus everything around it: the loop, the tool registry, the context management, the permission layer. Whether you assemble that scaffolding yourself from a framework or inherit it pre-wired from a harness like Claude Code or Cursor is the first architectural fork in any agent project.

Topics: Agent harness, frameworks, build-versus-adopt, open harnesses, vendor lock-in

Read →


The Day Copilot Started Charging by the Token: Anatomy of a Pricing-Model Break

June 2026

On April 27, 2026, GitHub announced that Copilot would abandon its flat-fee premium-request billing model for usage-based token pricing, naming multi-hour autonomous agentic sessions as the reason. The cleanest dated public instance of long-running agents breaking seat-based pricing.

Topics: Usage-based billing, seat-to-token pivot, GitHub Copilot, enterprise AI budgets

Read →


More Turns, Bigger Bill: Why Agentic Architecture Is a Token Multiplier Independent of Price

June 2026

The agentic cost spike isn't mainly a per-token-price story. It's an architecture story: autonomous loops take many more turns per task, so the same task costs multiples more even at a flat rate. The empirical counterpoint to "just route to a cheaper model."

Topics: Token multiplier, agentic turns, cost decomposition, price deflation, cadence and horizon

Read →


FinOps for Agents: A Practitioner Playbook for Routing, Gateways, and Prompt Caching

June 2026

Long-running agents turn AI spend into a variable cost that scales with autonomy, not headcount. A practitioner playbook for the four cost-control levers a platform team can deploy this quarter: spend caps, prompt caching, LLM gateways, and tiered routing.

Topics: FinOps, spend caps, prompt caching, LLM gateways, tiered routing

Read →


The Security Debt of Always-On Agents

April 2026

Why the enterprise security stack collapses when AI agents interact with data directly, and what a data-first defensive architecture for persistent agents looks like in practice. Covers identity-bound delegation, harness-level governance, and the tool/data enforcement points that bound persistent execution.

Topics: Persistent agents, identity-bound delegation, harness governance, tool gateways, data-layer controls, OpenClaw

Read →


From Copilot to Principal: How Always-On Agents Reorganize Knowledge Work

April 2026

Why persistent always-on AI agents do not shrink the human role, they invert it, and why most organizations are not structurally ready for the principal role this creates. Covers the overnight test, judgment bandwidth as the new bottleneck, and why delegation under governance is the next productivity frontier.

Topics: Principal-agent model, delegation under governance, coverage over speed, judgment bandwidth, enterprise operating model

Read →


The Specification Layer: Why Enterprises Can't Scale AI Development Without It

February 2026

Why explicit, machine-readable specifications are the missing infrastructure for scaling agentic development across enterprise teams. Covers AGENTS.md, CLAUDE.md, SDD frameworks, and the four-tier specification architecture.

Topics: Specification layer, AGENTS.md, CLAUDE.md, SDD frameworks, enterprise scaling

Read →


The Autonomous Agents Loop: Why AI Agents Produce Better Output When You Stop Interrupting Them

February 2026

Why autonomous AI execution loops outperform interactive assistance, and how enterprises can build the execution environment, context management, and multi-agent infrastructure to capture those gains.

Topics: Autonomous execution, Ralph technique, multi-agent architecture, context management, Plan-Execute-Verify-Iterate

Read →


About

These articles are companion pieces to the research papers, translating empirical findings into actionable guidance for engineering leaders and practitioners. They are part of an ongoing series on scaling agentic development for enterprise teams.

Looking for the papers?

The research papers these articles support live in Papers.

Updated at:

Released under the MIT License.