Context Engineering

Anthropic described how Claude uses offline compute, called “dreaming,” to organize memory. The process corrects memory mistakes, creates memories and builds representations for later retrieval.

Zeeg says Junior assigns an agent to verify tool calls. Its task threads share prompts, memory access and capabilities, but keep separate transcripts with more aggressive compaction for long-lived threads.

Anthropic’s public beta lets clients continue cached conversations by sending a previous message ID. Anthropic reports 35% lower median time to first token on Claude Code requests with 500K–1M tokens, at standard token rates.

Zed's DeltaDB links edits to conversations in a Git-compatible causal graph. It supports offline synchronization and snapshot replay, with each thread synchronized through its own Cloudflare Durable Object.

Exa 3.0 adds cross-result highlights, a larger index and Lite search at $4 per thousand queries. Artificial Analysis finds dynamic highlights increase task latency compared with standard search.

Matt Pocock's pre-commit hook rejects staged files above roughly 5,000 tokens, with generated-file exclusions and split instructions. A related skill requires existing oversized files to shrink with each commit.

The open ORA harness combines robot descriptions, perception tools and file-backed context. It reports 69.2% on RoboTwin and 81.4% on LIBERO-Pro, alongside preliminary real-arm demos.

GPT-6.1 Sol now defaults to 500K tokens in Codex and ChatGPT Work, up from 272K, to reduce compactions. OpenAI says subscriptions avoid long-context fees, and users can override the default up to 1M tokens.

Asana reportedly reduced its GPT-6.1 Sol browser-agent cost from $1.97 to $0.47 per run with an 89% cache hit rate. The report describes append-only prompt history and a Codex cache-measurement command.

ChatGPT dots can start Codex tasks or follow up on threads using conversation and automation context. The update also adds dot creation in the mobile app and improved thread controls.

Cognition’s Memory and Dreaming features build a memory graph across Devin sessions and revise stale records overnight. The open-source system is designed to work across harnesses and environments.

Aleph Alpha released Kolibri, an Apache 2.0 open-weight MoE with a 1M-token context window. The model has 78B total parameters and 3.46B active parameters, plus reasoning, tool calling and German-focused training.

OpenAI says Dot retains context across apps, coordinates Codex tasks, and flags work needing attention. Practitioner reports describe using it to reduce the overhead of managing coding agents.

Perplexity released pplx-embed-v2-context-9b-preview, which encodes chunks using whole-document context. Perplexity reports leading results on ConTEB and Turbopuffer's context benchmark.

A case study reports reducing a roughly 30,000-token AGENTS.md file by almost 50% while improving instruction quality. Session transcripts informed which rules remained and which moved into a skill.

OpenAI released GPT-6 Sol and GPT-6 Luna at API prices roughly half those of their GPT-5.6 predecessors. The models add controllable prompt-cache breakpoints and are rolling out in Codex and ChatGPT Work.

Teknium’s public evaluation says a Jev compaction strategy removes tool calls and eventually stops yielding savings. Repeated compaction can invalidate caches and increase total token costs, according to the critique.

Tencent released Hy4 Preview, a 770B-parameter mixture-of-experts model with 49B active parameters and a 1M-token context window. vLLM added day-zero support, while Cline, OpenCode Go, and Vercel AI Gateway made the model available.

The Prefix Sliding paper introduces an inference method that preserves the task prefix and a recent-token window while discarding older reasoning tokens. Its authors report up to 3× faster inference without retraining and longer reinforcement-learning rollouts.

Glean says its runtime routes enterprise-agent work across more than 40 models using company context. It reports $0.58 per query and 78% user preference over Claude Cowork in a 180-person benchmark.

A study of 557 coding-agent sessions finds instruction files and working notes account for most of what agents read. Related work puts installed skills' standing prompt cost at 50–280 tokens, while Backpass turns past sessions into reviewable AGENTS.M files.

A University of Pennsylvania study found common compactors retained only 17% of standing session rules while preserving task information. Practitioners use verified Markdown handoffs and background compaction to manage long coding-agent runs.

Codex now lets ChatGPT subscribers enable a 1 million-token context window for GPT-5.6 Sol. Automatic compaction starts at 900,000 tokens, and tokens beyond the default window count double against limits.

A practitioner report found speculative-decoding acceptance fell from 0.71 to 0.18 beyond about 32K context. Separate DSpark and mlx-dspark tests reported speedups on RTX and Apple Silicon setups.

The Echo Gap paper found self-improving agents can store wrongly self-scored episodes. Tested models endorsed 31% to 54% of their own wrong answers, while other work proposed RL-trained harness state and in-model memory.

Practitioner threads argued coding agents work better with task-specific harnesses, compact runbooks, evals, and deterministic checks. Examples included Sentry MCP traces and state-machine guardrails.

Practitioner posts argued agent evals should check final world state and tool-call trajectories, not just single outputs. A 288-run AGENTS.md test found context files did not improve correctness.

A paper summary found agent skills caused regressions across nearly 6,000 paired office-automation runs. VS Code also added prompt-to-skill migration, making the workflow more accessible despite reliability cautions.

A UW study found coding agents can refuse malicious memory-file instructions while preserving the payload for later use. The state-management debate also covered callable memories, reusable skills, and session-history eval pipelines.

Anthropic said Claude Code cut more than 80% of its system prompt without hurting coding evals. Other workflows use filesystem research folders, session IDs, wiki memory, and composable memory APIs to keep context durable.

Practitioners are sharing smaller, staged coding-agent loops instead of one large prompt. Examples include a Claude Code software factory, project-specific review agents, searchable logs, and harness layers for evals and observability.

Firecrawl introduced a new /search system that returns relevant excerpts instead of full pages for agent workflows. The launch claims 94.7% on SimpleQA and roughly 10x fewer tokens than full-page processing.

Developers described agents writing learned context into linked Markdown pages instead of repeatedly rebuilding it from RAG chunks. Related posts argued Markdown docs and llms.txt can improve agent navigation.

Baseten reports that injected facts can retain log-probability lift in weights while disappearing from answers, with prompts recovering 77–80% accuracy. OpenWiki adopted Open Knowledge Format for portable memory files.

Practitioners reported coding agents following outdated agents.md files and generic project instructions. Other posts tied long-context drift and weak domain language to cleanup-heavy coding runs.

Matt Pocock argued teams should delete docs that restate code and keep ADRs, glossaries, and thin navigation pointers. Zeeg reported colocated repo specs drifted and did not steer agents enough.

Practitioners described GitHub or folder-based markdown knowledge bases that feed persistent company or personal context to Codex, Claude Code, and Hermes. OpenWiki added codebase and personal brain modes for the same pattern.

One Claude Code user parsed 95 sessions and estimated subagent prompt caching made total spend about 8% high. pxpipe separately rendered dense text as images to cut context cost, with exactness tradeoffs.

Sakana listed 11 ICML 2026 papers covering LLM speed, memory, and agent evaluation. The lineup includes TwELL sparse kernels, RePo context positioning, CoffeeBench long-horizon agents, SoftMatcha 2 corpus search, Doc-to-LoRA memory, and Fast-weight Product Key Memory.

Condense.chat opened a compression proxy that strips tokens with Helene 1 and compacts settled agent loops with Adeline 1 to about 9% of their size. The service claims 100M saved tokens and 3× plan extension for Claude or Codex users, so test it on non-sensitive workflows first.

A day after /goal and thread automations landed in Codex, practitioners started standardizing on /goal specs, /fork or /side detours, and /rewind plus /compact recovery. The pattern matters because verifier design and compaction timing now control how well long runs hold together.

Junior’s first memory system cut one analytics task from about 3 minutes to 1 minute in early tests, with tokens down two-thirds and tool calls down 60%. The feature moves persistent task learning into the agent loop, though the results are still internal.

Builders released reusable loop artifacts this week, including a Loop Library Skill, repo templates, and published control-loop definitions for docs sweeps, onboarding checks, and error triage. It matters because teams are turning one-shot prompting into persistent agent runs with explicit stop conditions and shared repo state.

Plannotator 0.21.0 added direct document editing, an embedded CLI agent for co-context, first-class HTML handling, and Bedrock and Vertex support. It matters because the tool is moving from annotation-only review toward persistent IDE-like agent sessions with shared file diffs and cloud model backends.

Perplexity rolled out Brain, a self-updating context graph that carries prior sessions, files, and decisions into new Computer tasks. In research preview for Max users, it matters because Perplexity says the memory layer improves correctness and recall while lowering per-task cost on history-dependent work.

Two days after Factory 2.0 launched, Factory added AutoWiki, which turns a repo into structured engineering docs that update on every push and can be generated from Droid with /wiki. The feature keeps architecture and request-path context aligned with the live codebase.

Z.ai released GLM-5.2 MIT-licensed open weights with 1M context and broad runtime support. Vendor and arena results put it near frontier closed models on long-horizon coding.

Codex users are having the agent write its own `/goal` and sub-agent goals, with OpenAI-side commentary describing that as a built-in meta-prompting pattern. The workflow turns long autonomous runs into a tighter control loop, but users still review goals first so a bad objective does not burn tokens for hours.

Users are using Fable 5 as a planner and long-run orchestrator while pushing implementation and heavy reasoning to Opus and Codex. The setup keeps Fable on supervision and planning, so teams can track execution through live status pages on larger tasks.

Hyperbrowser shipped a Claude Code harness, InsForge showed a Fable run drop from 5.5M to 2.3M tokens, and Higgsfield published new MCP workflows. These tools add reusable harness, context, and interface layers around Fable for more controlled runs.