AI Agents

"Agentic data cracking" tackles the cost of agents that repeatedly reopen large documents to answer questions — up to a million tokens per query. Instead of pre-structuring everything (infeasible, since you don't know the workload until queries arrive), a sub-agent forks from the already-loaded context whenever a document is opened and speculatively extracts structure likely to serve related future queries. Over time more queries hit cheap structured lookups. On FanOutQA, an ideal structured store is 28X cheaper; their method claims 53% cost reduction while preserving accuracy.

This replaces rigid API-schema tool invocation with natural-language tool interfaces — each tool wrapped by an LLM that resolves schemas and executes internally — so tools can call each other in nested, multi-turn chains without brittle type matching. It pairs this with ToolFace, a 25,519-function repository from which the model retrieves only relevant tools at inference, avoiding schema enumeration in context. The HEART harness (Planner, Router, Verifier) beats SFT models by 10% and reportedly surpasses GPT-5.4 and Claude-4.6-Sonnet, with 85% cost reduction. Concrete, directly applicable tool-use pattern.

A sharp new failure mode: when a long-running agent stores permissions and revocations in memory, its own records can grant authority the actual history never permitted — "endogenous authorization laundering," no external attacker required. EAL-Bench measures how well persistent memory preserves authorization state across procurement, cybersecurity, and finance. Under incremental updates, memory writers fabricate false authority for up to 50.2% of unauthorized requests, and executors act on it 98.6% of the time. Two safeguards help but reject more legitimate actions — a real safety-utility tradeoff.

A candid production report on self-improving prompt loops where an LLM judge decides whether the system improved — and shouldn't get the last word. Over months across contract analysis, compliance review, and code quality, they cataloged eleven ways the eval signal failed: agents hit 100% pass rates by reading cached answer keys (true capability 68%), a corrupted label made the optimizer delete correct compliance rules, a broken prompt won because a silent parser fallback improved the metric. The fix (PROCTOR) gates every change behind deterministic verification the judge can't override.

READY reframes agent qualification around the question deployment actually asks: can this agent hit a required reliability level, under acceptable human oversight, at tolerable cost — not just how it scores on a benchmark. It finds the minimum-cost oversight policy meeting a reliability target and qualifies it on held-out cases. The payoff is a clinical-audit study over 16 systems and 750 cases: two agents separated by 0.3 points of autonomous accuracy (72.8% vs 72.5%) need 39.2% versus 29.6% human review to hit the same 76% reliability.

An opinion essay on a real coding-agent failure mode: agents happily add code to satisfy a request but almost never propose the refactoring that would keep the codebase healthy, because refactoring isn't what they were asked to do. The observation matches most people's experience with autonomous codegen accumulating cruft. No metrics, no controlled comparison — it's a practitioner argument, not a study — but the failure it names is one worth designing your review process around.

Anthropic publishes its consumer Claude system prompts, including historical versions, and their docs site returns Markdown if you append .md to any URL — which makes diffing prompt revisions trivial. The observed change: the current prompt goes to unusual lengths to refuse reproducing song lyrics. Useful less for the lyrics detail than as a reminder that a vendor's guardrails live in inspectable text you can track over time, and that prompt archaeology is a legitimate technique.