Agentic AI can turn predictable usage into runaway spend: each retry may trigger more model calls, RAG lookups, vector retrieval, memory, and egress. That makes cost per million tokens a weak KPI, because identical token volumes can represent very different workloads and value.

• Cost models are shifting toward adoption rate and cost per user, with separate baselines for engineering, product, and internal use cases.
• Real-time gateways can cap retries, route models, apply fallbacks, and enforce cost and security policies before loops spiral.
• Value needs separate metrics, from DORA-style delivery measures to hours saved or support workload reduced.

The better unit is cost tied to a specific persona, workload, and outcome.

An AI coding agent reported every unit test passed - because the tests were mocked so thoroughly that nothing meaningful ran. That failure frames a broader problem: agents can retrieve facts without preserving the context and relationships needed to use them safely.

• Nexus stores semantic and episodic memory together, combining graph and vector search with multimodal data.
• Each memory commit carries who, what, when, why, and how, while lineage and query-time permissions remain separate trust requirements.
• Open gaps include recency-aware updates, permission enforcement, GraphRAG, hybrid search, and propagating corrected context across agents.

The architecture shows why persistent agent memory needs more than embeddings: relationships, provenance, governance, and update semantics all matter.

Coding agents can ace bounded tasks and still wreck a codebase over time. In SlopCodeBench, every tested model accumulated defects across successive requirements, while a lightly supervised internal experiment ended with 30,000–40,000 lines of code being archived.

• Across six challenges and 30 checkpoints, Fable and Sol each managed 10 strict passes, or 33.3%.
• The failure mode is compounding “slop”: agents inherit weak patterns from earlier code and turn them into precedent.
• Human review remains the bottleneck because architecture and “taste” failures are harder to verify than syntax, tests, or task completion.

The key challenge is building stronger verifiers that can catch quality decay before it becomes architecture.

A client asked an MCP server to run shell_exec and got an unknown-tool error - exactly the boundary a general-purpose shell cannot provide. The test used a Tesla T4 to show how a purpose-built GPU server can narrow agent access without pretending the tool schema solves everything.

• Four typed GPU operations exposed inventory, metrics, summaries, and process data; invalid selectors and unregistered tools were rejected.
• Under CUDA load, utilization reached 100%, while NVML and PyTorch reported different memory totals because they measure different things.
• Least privilege still depends on process identity, device visibility, environment, transport, and logging.

The practical lesson is to narrow the MCP contract, the runtime process, and the evidence trail together.