Semantic Content Does the Text-to-SQL Work; Scaffolding Adds 0 to 15 Points

Empty every prose field of a data contract, hold the tool surface, retrieval instruction, table allow-list and operation rules byte-for-byte fixed, and hard-task text-to-SQL accuracy falls back to within 0 to 15 points of having no context layer at all. That is the central result of Qing Ye’s four-arm ablation on DABStep, posted to cs.DB on September 22, and it answers a question with/without comparisons have dodged: when a curated context layer lifts an analytics agent, is it the semantics, the retrieval scaffolding that delivers them, or the pre-computed views that ride along? Content dominates. Across four models the contract arm raises hard-task accuracy from 13.9% to 55.1%, 22.6% to 56.6%, 22.9% to 68.4% and 37.0% to 77.4%, and beats the same knowledge pasted into the prompt on every model. Scaffolding without content is worth 0 to 5 points on two flash models and 14 to 15 on two frontier models. The mechanism is lookup distance: the contract’s SQL carries the domain’s fee semantics up to 98% of the time, against 4% for the prompt arm, because the rule an agent needs sits one retrieval away instead of buried in a long prompt, and it does so at the lowest cost per correct answer on three of four models. Compiling the contract’s own SQL into views sets a ceiling (gold on all 176 covered tasks), and the contract arm’s shortfall against that ceiling shrinks from 39 points to 5 as model capability rises, so pre-computed macros earn their place only where an agent demonstrably fails to derive. Two governance details matter as much as the accuracy: ungoverned arms submitted 166 mutating statements, governed arms none, and the gain is confined to the contract’s domain. The paper’s own prescription is semantics first, scaffolding second, macros where derivation fails.

A second ablation, from a different corner, lands on the same conclusion about specifications. Sabina-Cristiana Necula’s Eurostat benchmark runs 30 natural-language tasks over seven datasets and four difficulty tiers through Claude Sonnet 5 in three replicates, 360 task-runs in all, under four conditions: task only, task plus a frozen metadata card, metadata plus a repair loop driven by sanitized execution feedback, and metadata plus the same attempt budget with no diagnostics at all. Exact correctness requires the right dataset, filters, output shape, values and unit, not merely code that runs. The headline is about the evaluator: a companion run under an under-specified output contract, where the ranking key and unit representation were never stated, understated the feedback condition by 23.4 points. Contract and evaluator design can dominate measured agent error, and the prescription for statistical coding agents is semantic validation against frozen specifications, a fully specified output contract, and a retry budget, with execution diagnostics contributing less than the budget itself.

AkasicMEM (Bae, Kim, Hur, Han and Kim, cs.DB, September 23) takes governance into the memory tier. When enterprise agents persist what they learn from governed sources and reuse it across tasks and principals, source restrictions can be bypassed through derivation: a row-restricted fact becomes a memory, the memory becomes another memory, and the second consumer never held the original permission. The paper names the property that closes the gap, authorization continuity, and realizes it through transitive lineage, policy composition at memory-formation time, and policy re-evaluation at retrieval time, on top of GraphAI’s AkasicDB, a unified vector, graph and relational store. It is a systems paper without accuracy numbers, but it is the first framing I have seen that treats source-memory integration, memory governance and authorization continuity as one design target across the memory lifecycle rather than three separate features.

On the orchestration side, the finding of the day is that parallel coding agents interfere far more on controlled tasks than mined pull requests would suggest. Xia, Wu and Park’s stale benchmark runs the same tests on each patch alone and on the merged pair, counting only failures the combination introduces. Among 834 runs on 417 mined Django PR pairs, a single run showed interference once the grading procedure was corrected. On constructed tasks built from 12 real Django helpers, where one agent changes an interface or rule another still relies on, interference occurred in 97% of runs, and a message describing the completed concurrent change recovered 82% of them. The authors are careful that the constructed rate does not estimate real-world frequency; reviewed PRs may simply contain few unresolved parallel changes because humans resolved them. The practical reading is that a cheap coordination channel, one sentence about what a sibling just changed, closes most of the gap when it does open.

Microsoft’s Agensh (Zhan, Song, Dong, Huang, Lian, Xia and Wei) removes the central orchestrator entirely. Workers run a cooperation loop over a shared workspace of proposed, ongoing and completed work, a message interface, and shared context that holds reusable findings and stated intentions; they claim sub-tasks, act, share, verify and merge asynchronously. On the five hardest ProgramBench tasks with GPT-5.6-sol (high), scaling from 1 to 128 agents raises the mean final test-pass rate from 19.31% to 28.78%, and on pandoc, 1 to 1,024 agents moves it from 33.89% to 55.06%. Larger organizations also reach a given pass rate earlier in wall-clock time. Read next to stale, the question worth asking is how much of Agensh’s gain comes from the shared-context component doing exactly the job stale’s one-line message did.

Two papers address the cost of knowing whether an agent got better. Ayyad, Wang, Shin, Zou and Adams attack SWE-agent regression testing, where each agent update can cost hundreds of millions of tokens to re-benchmark. They group instances by their outcome in the last full run to preserve the historical pass/fail rate, then select deterministically in trajectory-embedding space, picking instances nearest each group’s centroid. Across 76 configurations and three regression scenarios (same-config reruns, model or configuration changes, framework changes), a 10% trajectory-aware subset keeps median estimation error below 5% while cutting token cost by roughly 90%, and beats the strongest baseline’s 95th-percentile draw by 38 to 46% on worst-case error. The embedding of what the agent did turns out to be a better stratifier than what it scored.

And for retrieval, Wang et al. reframe agent-memory retrieval as budgeted evidence completion: not “is a relevant fragment in the top k” but “does the complete gold evidence set fit in a 2,048-token budget”. They build provenance units aligned to tool arguments and outputs and refine frozen dense scores with a zero-initialized residual R-GCN over typed provenance edges, evaluated on 2,000 span-grounded queries over 1,207 held-out ISETrace trajectories. Source-aligned units lift Full Support@2048 by 19.07 points over flat 512-token windows and stay 11.96 above a per-metric oracle over four flat chunk sizes. Graph propagation adds a smaller 4.55 points (95% CI 2.98 to 6.18), concentrated where gold evidence spans multiple events; entity co-occurrence expansion produces nothing comparable. Granularity is the dominant lever, and structure is a targeted second.

Three of today’s seven papers are, at bottom, about the same thing: a frozen, machine-readable specification (a data contract, a metadata card, a governed memory policy) doing work that prose in a prompt cannot, and measured error that shrinks once the specification is complete. Watch for whether the context-layer ablation replicates on BIRD or Spider 2.0, where the contract domain is less self-contained than DABStep’s, and whether anyone runs stale’s coordination message inside a harness like Agensh and measures how much of the thousand-agent gain it accounts for on its own.

Feed note: the copilot’s mirror_status reported a last sync of 2026-09-01, yet the items table carried entries timestamped through 2026-09-23. The freshness counter looks stale rather than the data; items were selected by their own published_at.