AI Agents

Context compaction keeps GPU memory constant for long agentic traces, but most strategies re-prefill the context repeatedly, killing training throughput. KV-streams streams the KV cache forward instead of flushing it after each compaction, yielding 2.6–5x wall-clock speedup across three compaction strategies with no measured performance loss. The more interesting side finding: the streamed cache acts as a recurrent state carrying information long gone from context, and RL alone is enough to elicit that behavior — no architectural change required.

Token consumption for the same agent task can vary by an order of magnitude across runs, because each tool result reshapes the plan and the growing context inflates every later call. TokenCast learns a composable per-segment cost representation that captures the re-read cost of earlier context, refreshing its forecast from observed evidence with no extra LLM calls (32.8ms mean per run on SWE-bench Verified). It averages 14.5% MAE reduction over the strongest baseline and, in replay, uses 21.3% fewer tokens than a fixed budget at matched completion.

Long-horizon runs turn execution control into its own task: which partial work to build on, when to restart, when to stop. This proposes a meta-reasoning harness where a controller consolidates run state, scores next options against remaining budget, and dispatches workers with memory-drawn context — carrying only a compact account between decisions rather than replaying history. On ProgramBench it hits 71.5% with GPT-5.5 vs 58.0% for Codex, gains 3.6–4.2 points elsewhere, and keeps improving where direct control plateaus. The overhead hurts at small budgets.

Agents mutate state across local files, tools, and remote services, and today's harnesses lack a unified way to undo it safely. Planarian introduces statepoints — consistent, restorable point-in-time versions spanning local and remote state. Snapshot uses incremental process and filesystem checkpointing locally while recording compensating actions for remote changes; rollback reverts and replays those compensations; fork branches isolated executions. The compensating-action model for services that don't support checkpoints is the clever part, since most external APIs won't ever expose one.

When a corpus is exposed as flat files, an agent rediscovers cross-document relationships on every query, missing complementary evidence and burning tokens. CorpusMap builds a navigation layer offline by resolving recurring entities across documents into Entity Pages that link to every mention, forming a traversable entity-document graph shared across all queries rather than re-derived per run. Evaluated on 7 models across 3 datasets, it improves evidence discovery while cutting tokens. The offline-construction, shared-links design is the reusable idea for multi-hop agentic search.

An empirical privacy study of conversational AI agents across web and mobile finds they behave like any other tracker-laden surface — telemetry, third-party beacons, and fingerprinting stitched into the assistant layer users assume is private. The HN thread (418 points) suggests this touched a nerve. If you're embedding an agent front-end, the finding is that the conversational channel inherits the same surveillance plumbing as the page it sits on, unless you actively strip it.

A practitioner post arguing against MCP as the default tool-integration layer, presumably favoring plainer function-calling or direct API wiring. Engagement is modest (74 points, 27 comments) and the linked content isn't excerpted here, so the depth of the argument and what it proposes instead aren't visible from the input. The MCP-skepticism genre is worth tracking as the protocol calcifies, but read the post before treating this as a settled critique.