Voices Digest — two weeks to 2026-07-12

The first entry in a recurring News-format digest of the voices we track (agentic-velocity T1–T3): what each shipped, said, or measured in the two weeks leading to 2026-07-12, with links. Closes with our own contributions this period.

About this digest

First entry in a recurring, event-driven monthly series. We snapshot what the voices in Agentic Velocity — Voices to follow shipped, said, or measured, with links to primary sources. Each voice gets 1–3 bullets; anything longer earns a section on the relevant living paper. Closes with our own contributions and where we are looking for collaboration.

Voices — the two-week snapshot

Relace

• Published "A Year of Fast Apply": LoRA SFT on 3–8B code models + speculative decoding reaches ~10k tok/s at 98%+ merge accuracy; frontier models still fail diff formatting 4–10% without a merge layer. (source)
• Shipped Fast Agentic Search (FAS): RL-trained search subagent with 4–12 parallel tool calls/turn, 4× latency reduction at Claude-4.5-Sonnet accuracy; OpenAI-compatible endpoint, also on OpenRouter. (source)

Morph

• WarpGrep v2: SWE-Bench Pro #1 on multi-repo + package + log search; ships inside the same morphmcp server as Fast Apply. (source)
• Reflexes preview: sub-turn reflex primitives that batch mechanical follow-ups (verify, retry, re-diff) without a new LLM call — the "kill sequential-edit spam" pitch, at the harness layer. (source)

Cognition

• SWE-1.7 / Fusion: model family that folds planning and edit-format routing into one contract, anchored to the 60%-of-turns-on-search measurement Cognition established with SWE-grep. (source)

Agent Skills ecosystem

• Six-month audit: 40+ tools adopted the SKILL.md spec; skills.sh, SkillsMP, and VoltAgent/awesome-agent-skills are the three surviving marketplaces; official vendor skills exist from dbt Labs, Supabase, and Vercel. (source)
• SkillsBench proposal circulated: community benchmark that scores skills on activation rate, task-success delta with the skill active, and context overhead — the numbers that will decide which skills survive. (source)
• Snyk ToxicSkills audit: 13.4% of scanned marketplace skills had critical issues, with confirmed malicious payloads in the wild. No signing, no verified publishers yet. (source)

ast-grep

• ast-grep Outline: structural-outline mode emitting a file’s AST landmarks (functions, classes, imports) as a compact map — a zero-embedding alternative to Aider-style repomaps. (source)

Simon Willison

• Wrote up "agentic-engineering-patterns" — a durable pattern-language for coding-agent workflows: subagent contract, memory-tool routing, receipt-first debugging. (source)
• Argued "CLI beats MCP for coding agents": agent-facing CLIs (gh, git, wrangler, scoped shell) beat wrapper MCPs on latency, parallelism, and legibility — informs the workspace AGENTS routing table. (source)

Cursor engineering

• Secure Codebase Indexing writeup: content-hash-keyed cache + Merkle diff + simhash worktree reuse — the technique we cite in memory-tooling D2 for the freshness architecture upgrade. (source)
• "Reward hacking is swamping model intelligence gains": warning shot on RLHF-only scaling. Argues for deterministic tools + typed contracts + validate gates around the model — our stack. (source)

Sourcegraph + Amp

• Agentic Batch Changes public beta: cross-repo migrations orchestrated by agents on top of SCIP; the closest external analog to our wave methodology. Worth studying for the pattern differences. (source)
• "MCP + a cheaper model beats a bigger model alone" (2026-06-16): concrete evidence retrieval quality dominates model choice at fixed cost. Reinforces our tooling-health-outranks-feature-work rule. (source)

Anthropic Engineering

• Harness posts (Sean Young, Neal Rajasekaran): first-party writeups on Claude Code harness architecture, tool contract choices, and edit-format receipts. (source)
• Advanced tool use / Programmatic Tool Calling: batch N tool calls in a client-executed code block instead of one-per-turn. Free win against sequential-edit spam wherever the harness supports it. (source)
• Skills spec 6-month retrospective: fastest cross-vendor standardisation in AI tooling; the Skills body is now a first-class distribution channel for portable procedural knowledge. (source)

Voyage AI

• voyage-context-4 (2026-06-29): chunk-context-aware embeddings that ingest surrounding chunks at embedding time — the "stop worrying about chunking" pitch. We adopted it under memory-tooling D6 as option (b) newest-first-evidence-checked. (source)

Letta

• Context Repositories: git-backed memory framework for canonical-doc management. We rejected Letta as an agent memory system, but the doc-management framing is a distinct question — reopened as memory-tooling D12. (source)

Addy Osmani

• Agent-skills authoring craft: gotcha lists beat generic advice; plan-validate-execute loops for batch/destructive ops; bundle deterministic scripts when the agent keeps reinventing logic; skills for portable procedure, AGENTS.md for repo context. (source)

Our contributions this period

• 277 GB / 412 GB / 202 GB runaway-DB incident series (Incidents #1–#3) — Three runaway ChunkHound DBs inside four days; ~863 GB reclaimed. Root causes documented: DuckDB HNSW drop-recreate blocks ART/row-group cleanup (duckdb#19468); relative database.path bypassed the workspace cap; the CLI bulk-index path was a cap-bypass with SUCCESS reporting despite blowing the cap ~10×.
• chunkhound#365 — CLI bulk path ignores max_disk_usage_mb — Upstream issue filed after Incident #3: the disk-cap guard lives on the daemon write path only; the CLI bulk indexer takes a different path that does not consult the guard, and reports SUCCESS after blowing the cap. Triggered D4’s reopening.
• codesearch#148 — Contribution / issue against codesearch. codesearch was healthy through all three incidents and remains the reference for the HTTP-behind-supervisor MCP transport pattern (D5).
• MAB-over-harnesses methodology — First published methodology for a live bandit over agentic coding harnesses on identical KYM agent cards with receipt-anchored dual evaluation. Runner: nexartis-parallel-waves-agent-harness (renamed 2026-07-12 from nexartis-opencode-parallel-waves per memory-tooling D10). See the methodology paper for the full slate.

Where we are looking for collaboration

Ozzydev is the on-machine agentic client — reference implementation of the local-mac Cubicle runtime host — and we are actively looking for contribution and commercial partnerships around the surfaces this digest touches: harnesses, memory stack, skills ecosystem, and canonical-doc management. If any of the work above is your work, or your product, or your workspace, we would like to talk.

Cubicube is the packaged form: dev-team-in-a-cubicle, cloud-deployable, built on the same substrate the MCP controller/manager (memory-tooling D5, expanded scope) is now the foundation for. Commercial pilots and design-partner conversations are open.

How this digest stays current

Roughly monthly, event-driven, timestamped. Each edition is a new page at /research/voices-digest-YYYY-MM; existing editions are frozen (status: published), not updated in place. Errata get their own bullet in the following edition. The index carries the current edition; older editions link forward from their footer as they land.