# An agent should not have to read the whole team Broadcasting everything burdens every agent. Persistent, searchable contributions can help participants reuse work while mutation safety remains a separate concern. _Architecture · 2026-10-19 · 4 min read · by Ablo Team_ As more participants contribute, the shared space needs a way to make useful work discoverable without delivering every update to everyone. ## Key takeaways - Let participants retrieve relevant work instead of broadcasting all context. - Preserve useful contributions after their producer stops. - Separate discovery of information from authority to mutate shared resources. A research team grows from two agents to twenty. Each sends its findings to every other agent. Soon, each participant spends more effort reading progress updates than investigating its own question. The team has shared context, but attention has become the bottleneck. A different reading path lets an agent ask what has already been found, retrieve a small relevant contribution, and follow its evidence when needed. The shared environment becomes a place to pick up work rather than a transcript everyone must consume. ## Publish something another participant can use A useful contribution states the work attempted, the result, the evidence behind it, and any remaining uncertainty. It should be possible to reuse the result without reconstructing the producer’s entire conversation. It should also be possible to inspect the source when that result matters to a new decision. ![Creatures retrieve selected useful components from a shared workspace instead of carrying every object to every participant.](/blog/coordination/shared-discovery.webp) Shared work becomes useful when a participant can find the relevant contribution and inspect what it rests on. ScaleMAS explores this with an append-only message datastore and publish, search, and visit operations. Search combines lexical relevance with citation-based weighting. Reads return a bounded number of bounded-length messages; that bounds each read, not the whole run’s context or compute. The datastore persists during the task rather than serving as permanent cross-session memory. ## Reuse is a signal, not a guarantee of truth If several useful contributions build on an earlier result, that result may be worth surfacing. It can also be repeatedly cited because it is convenient or confidently phrased. A retrieval ranking should help a participant find evidence; it should not turn popularity into authorization or factual certainty. Keeping references inspectable makes this distinction practical. A later participant can examine the original source, notice a scope limit, or publish a correction. The workflow also needs a clear way to represent superseded conclusions, especially if older contributions remain available for search. ## Interpret scaling results with their scoring rules The ScaleMAS study extends to 128 agents on selected tasks. Its third-party evaluations use Pass@N, which asks whether the group found a correct answer rather than whether it reliably selected one final answer. Its contradiction-audit score uses relaxed recall without penalizing extra reports. These choices limit what the headline results establish about production decisions. A searchable shared space can improve discovery while leaving a write unsafe. An agent that retrieves the right customer agreement still needs current authority and valid premises before changing an order. Finding a contribution, trusting its evidence, and committing a dependent action are different boundaries. ## Make continuity concrete For Lobby, the practical question is whether another participant can continue from a task’s recorded attempts, checks, and unresolved work after the original agent stops. For Ablo, it is how that shared work connects to the permissions and state checks of the integrated application. Searchable context and guarded mutation can complement one another without being the same feature. We would test this by stopping a contributor after it publishes a partial result, asking a new participant to find and use it, and checking the resulting action. Measure duplicated work, relevant evidence retrieved, final correctness, and total cost. A large store or a large agent count is not, by itself, evidence of useful continuity. ## Sources and further reading - [Scaling Multi-Agent Intelligence with Substrate-Centric Coordination](https://arxiv.org/abs/2610.16120) — Qi et al., October 2026 preprint. Full 29-page PDF reviewed; Sections 2–3 and Appendices A and C describe task-local persistence, bounded reads, evaluation subsets, and Pass@N scoring. ## Frequently asked questions ### Does a shared datastore make concurrent updates safe? No. It can support persistence and discovery. Mutating a shared resource still requires the authority, dependency validation, and consistency rules appropriate to that application. --- Canonical URL: https://www.abloatai.com/blog/shared-work-should-be-findable Also machine-readable on this domain: /llms.txt, /llms-full.txt, /openapi.json, /sitemap.xml, /robots.txt