Code generation has a free, instant, perfect oracle. Prose has none. Here is two months of narration-quality fixes in PaperCast, why each one was locally right and globally insufficient, and what we are changing.
Deploying Qwen3.8-27B on a DGX Spark shows why dense-model decode is bandwidth-bound, and how MTP speculative decoding lifts throughput from 11 to 29 tok/s.
MCP standardizes how tools and context are exposed to models, but reliable agent systems still depend on orchestration, policy, durability, observability, and approval-driven runtime design.
A practical reference blueprint for an agent harness, grounded in the live harness-engineering repo and updated with a real blueprint export instead of an imaginary architecture diagram.
Why agent performance engineering lives in the harness, with a real repo update that adds per-step timing and workload estimates to trace summaries and run summaries.
A practical look at what security actually means in an agent harness after adding a checked-in policy config, provider env handling, and verification surfaces to the live demo repo.
A practical look at small multi-agent systems after adding explicit planner→executor→reviewer handoffs to the live harness-engineering demo.
A practical look at harness-level sandboxing and isolation after adding explicit action categories, policy checks, and write-path enforcement to the live harness-engineering demo.
A practical look at tracing, operator-facing observability, and trace-aware evals in the live harness-engineering demo after adding persisted trace summaries and lightweight eval fixtures.
A practical look at human-in-the-loop design in the live harness-engineering demo: approval as explicit workflow state, not a chat courtesy, plus a new pending-action inspection surface for operators.
