Posts tagged "agents"
10 posts tagged with agents.
The /loop Pattern: When Self-Pacing Helps vs Burns Tokens
A practitioner's recipe for cache-aware long-running agents. When self-pacing wins, when it loses, and why the 270-second prompt cache window is the only number that matters for an idle loop.
Cheap Models, Honest Reports: Getting Haiku to Behave Like Sonnet
Haiku ran my fixture suite at parity with Sonnet once the doctrine forced verification and status tokens. The trick is not a better model. It is a tighter contract about what done means. This holds across both Anthropic and OpenAI's durable tier system.
Done Means Done: Why Status Tokens Beat Status Updates
Agents reporting completion in prose is the leading source of false-green builds. A four-token discipline (VERIFIED, UNVERIFIED, PENDING_REVIEW, FAIL) is what makes long-horizon agent work actually verifiable.
Frontier AI Performance at Home, Without the API Bill
Maestro fuses the model CLIs already on your machine into one answer that beats any of them alone. It is the same mixture-of-agents method OpenRouter's Fusion API benchmarked at frontier level for about half the cost, except it runs on the flat-rate subscriptions you already pay for instead of a metered API bill.
Why I Stopped Using Multi-Agent Frameworks (And What Replaced Them)
After 18 months of building with AutoGen, CrewAI, and LangGraph, I replaced all of them with explicit decomposition and deterministic orchestration. Here is why, and what to use instead.
Prompt Engineering Is Dead. Context Engineering Is the Job Now.
Clever prompt wording stopped being the bottleneck once models got good at following instructions. The real work moved to engineering what goes into the context window.
What 18 Months of Production AI Agents Actually Taught Me
Five things that broke, three that worked unexpectedly, and the cost arc that surprised me. A first-person retrospective from running AI agents in production for 18 months.
The Automation Paradox: Why AI Agents Create More Work Before They Create Less
Teams deploying AI agents see a productivity dip before gains materialize. The supervision tax, the correction loop, and the trust calibration period are predictable, manageable phases, not signs of failure.
Why Your Multi-Agent AI System Keeps Failing (And What the Research Actually Says)
Multi-agent LLM systems fail 41-87% of the time. 79% of those failures come from coordination, not capability. Here is what the research says about building systems that actually work.
AI Agents Need Governance, Not Guardrails
For practitioners and engineering leads evaluating governance approaches: SDK wrappers are in-process guardrails agents can bypass. Infrastructure-level governance, where agents never hold API keys, is the only enforceable approach. Here's why.