EXPERIMENTAL RESULTS

Research questions, findings and cases

Explore the four research questions, seven findings and their archived execution evidence.

A valid message can still carry incomplete evidence.

A valid but incomplete message is accepted as verified, while an independent evidence audit fails. — Figure 1. A communication fault can preserve a valid message while removing required evidence. The workflow reports success, but the independent evidence audit fails. Click to enlarge.

A missing field can leave a message syntactically valid. The receiving agent accepts it, the workflow continues, and the final answer lacks the evidence required to justify it.

What we study

We investigate how communication faults become visible to agents, propagate through multi-agent workflows, and affect task outcomes. Controlled fault injection across five task domains, three workflow topologies and three deployed model configurations reveals when information remains recoverable and which protection mechanisms help.

Explore results and cases →Read the methodology →

From communication exposure to task outcome

Figure 2. Conditional cross-layer propagation: lower-stack sources, receiver-visible symptoms, MAS consequences and independent task outcomes. Click to enlarge.

Figure 3. Sequential, Flat and Hierarchical workflows with request (I2), observation (I3) and evidence-handoff (I4) boundaries. Click to enlarge.

RQ1 · FAULT MANIFESTATION

How do communication faults propagate and remain undetected?

Figure 4. Cross-layer manifestation and failure visibility. Source mapping and silent-failure rates have separate denominators. Click to enlarge.

Figure 5. A partial-content message remains processable while required evidence is missing. Click to enlarge.

RQ2 · CONTEXT SENSITIVITY

When can affected information still be recovered?

Figure 6. Matched request loss at I2 and evidence-handoff loss at I4. Click to enlarge.

Figure 7. Task-failure rates under semantic corruption and partial-content delivery across five Flat task contexts. Click to enlarge.

RQ3 · TOPOLOGY AND MODEL

Which changes rescue the same faulted tasks?

Figure 8. Clean-conditional loss by topology/model and paired gains from switching to Flat. Click to enlarge.

Figure 9. Matched I4 non-delivery: Sequential and Hierarchical lose their evidence; Flat retains an intact direct delivery. Click to enlarge.

RQ4 · PROTECTION MECHANISMS

Which dependency should a protection mechanism repair?

Figure 10. Guarded evidence recovery repairs incomplete evidence without repeating the completed action. Click to enlarge.

Success rates under the main protection study: execution-side and evidence protection recover complementary failures. — Figure 11. Success rates under the main protection study. Execution-side and evidence protection recover complementary failures. Click to enlarge.

In the separate Reddit recovery study, protected partial-content and corruption runs both reach 77.8% success, matching the 77.8% clean baseline. Guarded readback also saves two HTTP calls on clean Shopping trials compared with always-on readback.

Explore the research artifact

Historical run explorer

Inspect archived messages and outcomes with stable task identities.

Methods and definitions

Review taxonomy sources, interception boundaries, exposure checks and denominators.

Read extended methodology →

Reproducible analyses

Paired data, analysis code and completed protection-study summaries.

Browse research artifact ↗

Scope & provenance

Frozen archives · source digests · explicit limits

Research findings and evidence

The study addresses four research questions and reports seven findings. Aggregate statistics summarize the study results; individual archived runs provide case-level evidence.

The replay catalog covers the common matrix, source-mapping fixtures and historical protection collections. Position-extension and later protection analyses are distributed separately with their own denominators.

What this artifact reproduces

This website replays archived evidence. It does not rerun an agent, execute a fault, or contact a model provider. Source identifiers link every record to an archived input file.

Task numbering

Task 001–100 identifies the 100 main-study tasks across models, topologies and repetitions. Original benchmark IDs remain visible. Shopping uses 001–030, Admin 031–050, Reddit 051–065, SWE-bench 066–085, and TheAgentCompany 086–100. The Reddit extension reuses these IDs; the separate protection study uses RQ4 Task IDs, and lower-stack tests use Fixture IDs. These numbers identify tasks or fixtures, not individual executions.

Download fixed task-ID mapping ↓

Reading the outcomes

Composite success follows the archived adapter endpoint; it can require an accepted evidence contract as well as a correct answer. Handling response is a recorded flag, not proof of a complete recovery chain.

Repeated executions are not controlled model sampling seeds. A missing consequence label is not proof that no propagation occurred.

Evidence availability

All verified records are indexed. Detailed exports cover representative condition/outcome groups, named case studies, all RQ4 runs, and fixture measurements. Other records expose only catalog fields and are labeled summary only.

URLs, credentials, email addresses and private deployment paths are redacted. Event lists preserve source order. Separate fault logs are not assigned to conversation steps without recorded links.

Collections

Important boundaries

Source-to-symptom fixtures have no MAS success score. Frozen-evidence SWE/TAC trials do not represent a fresh end-to-end benchmark execution for every injected fault. RQ4 is a separate Shopping mechanism study with finite exposure budgets.

Download manifest and SHA-256 digests ↓

From message faults to task outcomes