Agent reliability depends on context plumbing and verifiable handoffs

A large context window has limited value when every turn recomputes the full history or useful information becomes hard to retrieve. qMLX restores key-value (KV) attention state from disk and fixes prefix-matching and interrupted-history bugs. On one M3 Ultra, repeated prefill for 32,000 tokens fell from 88 seconds to 0.64 seconds; a short follow-up at 168,000 tokens reached its first token in 2.6 seconds.

ContextOps checks duplicated retrieval chunks, prompt growth, source concentration, and token imbalance before a large language model (LLM) runs. Its sample found 214 duplicated tokens and estimated 12% savings. A separate social-media excerpt reports retrieval falling from 80% at 256,000 tokens to 36% at one million, but gives no dataset or protocol. Together, these items support tighter context measurement; only the qMLX report provides detailed runtime measurements, and those come from a single machine.

AgentTransfer gives software identities, inboxes, storage, recipient-aware delivery, integrity checks, and signed receipts. Its offline demo transfers a 1 MiB file between two agents and verifies both the SHA-256 hash and receipt chain. The project also claims that its Model Context Protocol (MCP) bridge can stream files up to 5 GB without placing the bytes in model context, although it reports no throughput or reliability tests.

Mesh LLM pools trusted machines behind one OpenAI-compatible endpoint. Requests can run locally, route to a peer, or split model layers across several nodes. This makes distributed private inference accessible to existing clients, but the current evidence is functional: no latency, cost, or reliability benchmark is supplied.

Two field reports place coding agents across maintenance, design, implementation, execution, and review. Terence Tao used an agent to port roughly two dozen obsolete Java applets to JavaScript in a few hours, then built new mathematics visualizations. He found one minor migration bug, while the agent found two defects in the original code. The low-risk, supplemental nature of these tools made hands-on review an acceptable check.

A separate software-development guide uses agents to summarize issue threads, inspect related pull requests, discuss implementation options, run local commands, and produce file-level review notes. It offers no controlled measurements, but it shows a concrete division of labor: agents compress investigation and draft changes, while developers retain design and review decisions.