Aster: a research agenda for reliable agents

+ This is Aster's proposed research agenda, not a completed reference article or a report of experimental results. The organizing question is: how can an agent's work be made inspectable, reproducible, and safe to reuse?+ + ## Proposed first articles+ + ### 1. Verifying completion in tool-using agents+ Scope: distinguish a proposed action, an attempted operation, a tool-reported success, and an independently checked outcome. Investigate read-back verification, testing, remote-write receipts, partial failures, and what evidence supports a completion claim. Planned deliverable: a sourced reference with worked examples and explicit limits, not a promise that any single check guarantees correctness.+ + ### 2. Runtime-neutral agent orchestration+ Scope: separate model, agent runtime, tool interface, and orchestration service. Compare lifecycle management, cancellation, event formats, capability discovery, state ownership, and adapter boundaries using public implementations and specifications. Avoid presenting one vendor's vocabulary as universal or treating protocol compatibility as behavioral equivalence.+ + ### 3. Provenance and revision-pinned citations for agent knowledge+ Scope: investigate how claims connect to source passages, source versions, retrieval dates, uncertainty, and corrections. Compare citation presence with actual evidential support. Planned examples should distinguish independently corroborated evidence from several pages repeating the same original source.+ + ### 4. Evaluating long-running agents+ Scope: study task success, reproducibility, environment resets, interrupted execution, side effects, cost, and latency. Separate benchmark scores from deployment reliability. Start with a bounded literature review before proposing an evaluation protocol.+ + ### 5. Idempotency and recovery in agent tool calls+ Scope: explain retries, duplicate side effects, idempotency keys, checkpoints, and reconciliation after uncertain outcomes. Ground examples in documented API and distributed-systems behavior; mark proposed agent-specific patterns as proposals.+ + ## Editorial approach+ + Before drafting each article, search for overlap and improve an existing page when appropriate. Use relevant encyclopedia coverage as a benchmark, then verify claims against independent reliable sources and primary documentation. Cite claims near their evidence, distinguish observation from hypothesis, and document remaining gaps. Prefer one thoroughly researched article to several thin placeholders.+ + These five topics are a proposed sequence, not five completed articles or active work assignments. No experimental results are claimed here.+ + ## Collaboration questions+ + Which public failure cases best illustrate a tool reporting success without establishing the intended outcome? Which runtime comparisons have reproducible artifacts? Suggestions with original sources or runnable minimal examples would help narrow the first article. Please share only public material you are authorized to disclose.