Writing
Practical notes on run bundles, human gates, model-agnostic tracing, token economics, and AI governance for agents you operate in production.
Latest · LLM Workbench
When GPT-6 Astra runs for 27 minutes and can't show you its work, that's a governance failure. This week's news makes the case for run bundles, model-agnostic tracing, and human-in-the-loop gates.
· LLM Workbench
OpenAI's internal data shows coding agents reshaping research work and spend. Here's why that acceleration makes audit-ready run bundles, cost telemetry, and human-in-the-loop gates non-negotiable.
· LLM Workbench
Fable's premium pricing is pushing teams to route work across models by cost. Here's why model-agnostic tracing, cost telemetry, and audit-ready run bundles matter more than ever.
· LLM Workbench
A local 27B model that burns 22,000 reasoning tokens to draw a circle is a governance story. Here's why reasoning-effort defaults, model-agnostic tracing, and run bundles matter for auditable AI.
· LLM Workbench
A gym-booking exploit, retired model gateways, and cyber evals all point the same way this week: agentic AI needs audit-ready run bundles, human gates, and cost telemetry to stay governable.
· LLM Workbench
This week's managed agents, open-weight open letters, and a JSON-compaction release all point back to the same need: audit-ready run bundles, human gates, and model-agnostic tracing for AI you can defend.
· LLM Workbench
This week's news on token relay markets, prompt-injection-resistant models, and coding agents shows why audit-ready run bundles, cost telemetry, and human-in-the-loop gates matter more than ever.
· LLM Workbench
This week's AI news—from OpenAI's ROI scorecard to Hugging Face's security disclosure—makes a strong case for evidence-first LLM operations. Here's how run bundles, HITL gates, and cost telemetry connect the dots.
· LLM Workbench
This week's AI news—DRI accountability, managed agents in Gemini, GPT-5.6 in Copilot, and shifting model pricing—maps directly onto run bundles, human gates, and governance.
· LLM Workbench
A source-grounded look at how an agent-driven sqlite-utils release turns cost telemetry, cross-model review, and changelog provenance into a working model for audit-ready run bundles and governance.
· LLM Workbench
This week's open-weight agentic coder and a sharp critique of unreviewable PRs both point to the same need: audit-ready run bundles, human-controlled gates, and model-agnostic tracing.
· LLM Workbench
This week's agent-security news—OpenAI's Daybreak, Samsung's Codex rollout, and DeepMind's AI Control Roadmap—reframes why audit-ready run bundles and human-in-the-loop gates matter.
· LLM Workbench
This week's AI news—from disposable code to near-autonomous chemistry agents and deployment simulation—all points to the same operational gap: audit-ready run bundles, human gates, and model-agnostic tracing.
· LLM Workbench
This week's AI news—deployment simulation, multi-agent safety funding, and an open eval workbench—maps directly to run bundles, human gates, model-agnostic tracing, and cost telemetry.
· LLM Workbench
The runtime, the React shell, the AI SDK wrappers, and the MCP server are now published under @llm-workbench on npm, MIT-licensed. Here's what shipped, why, and the engineering it took to make it genuinely installable.
· LLM Workbench
A few weeks of work, recapped—removing the last unsafe-eval from production, a fail-closed audit gate, and a landing page that finally feels like the product.
· LLM Workbench
Ajv compiled JSON-Schema validators in the browser with new Function, which forced 'unsafe-eval' into our production policy. Here's how we precompiled it away—and pinned it shut.
· LLM Workbench
A demo should show the shape of a real agent run, not a toy. So ours runs beloved-story logistics—Gandalf's routing, a flux-capacitor power calc—on top of the actual engine.
· LLM Workbench
The run bundle is the unit LLM Workbench is built around. Here's what's inside one—and why each part earns its place.
· LLM Workbench
What if every person had a universal right to replay the AI decision that affected them—loan, hire, claim, refund? Here is the moonshot: tamper-evident run bundles as the substrate for adjudicable cognition.
· LLM Workbench
A recovering lawyer's case for a machine-readable contractual layer between AI agents and their human controllers—scope, consent, revocation, restitution—encoded as gates and run bundles.
· LLM Workbench
Hyperscalers are racing to subsidize LLM tokens, lock workloads in, and raise prices later. Here is how the playbook works—and how a model-agnostic control plane keeps you portable.
· LLM Workbench
A unified operating view for AI leaders—token-aware routing, context discipline, and tamper-evident run bundles as the control plane for trustworthy scale.
· LLM Workbench
Long contexts drift, tools lie politely, and black-box agents burn tokens on retries—here is why run-shaped visibility and bundles matter for finance-grade LLM operations.
· LLM Workbench
How tokenization shapes bills, why small fast models work for formatting and classification, and when you still need frontier reasoning—plus how to prove routing decisions in production.
· LLM Workbench
Roles that get leverage from bundles and gates today, plus a practical entry path through docs, traced calls, playgrounds, and machine-readable surfaces.
· LLM Workbench
From opaque prompts to replayable bundles: gates, traced model I/O, artifacts, integrity—and surfaces both humans and tools can consume.
· LLM Workbench
A reference plane for deployed agents—not only faster models, but runs you can replay, approve, audit, and hand to another engineer without folklore.