Anthropic's Dynamic Workflows announcement last month described hundreds of Claude agents, working in parallel, porting roughly 750,000 lines of Zig to Rust for the Bun runtime. Up to 1,000 subagents per run. Sixteen concurrent. A sandboxed orchestrator coordinating the whole thing. Eleven days from first commit to merge.
The scale is striking. The workflow's structure is more revealing. Hundreds of agents wrote Rust files in parallel, two reviewer agents checked each file, and a fix loop drove the build and test suite until both ran clean. The reviewers caught execution errors. The test suite caught everything else, and it's the test suite that made the 99.8% pass rate meaningful. That number depended entirely on an external, deterministic standard that existed before any agent touched the codebase. The agents agreeing with each other would have told you nothing.
The broader Dynamic Workflows architecture leans on a different verification model for domains without test suites. Independent agents adversarially review each other's findings. Multiple drafts get compared. Sources get cross-checked. The word "converge" appears. For code, convergence means passing tests. For everything else, convergence means multiple instances of the same base model, interpreting the same specification, arriving at the same answer.
A preprint by Aman Mehta studying Claude's failure patterns across repeated runs found that 71% of failures were consistent wrong interpretation. The same incorrect assumption, confidently applied, every time. Small study, single author, not yet peer-reviewed. But the mechanism it describes has a particular implication for parallel orchestration. If one agent consistently misinterprets a specification, a thousand agents inheriting that specification produce internally coherent, systematically mistaken output. The adversarial reviewers, sharing the same model and the same interpretive priors, are likely to agree that the wrong answer looks right.
One way to read Anthropic's hard limits is as an implicit theory of anticipated failure. The 1,000-agent cap prevents runaway loops. The sandboxed orchestrator bans non-deterministic built-ins, can't touch the filesystem, requires plain-literal metadata. Each constraint implies a failure mode someone thought through carefully. The MAST failure taxonomy, validated across 1,600+ multi-agent execution traces, found that specification problems account for 42% of breakdowns. What's notably absent from the constraint set is anything that detects whether the specification itself was correctly understood before hundreds of agents start executing against it.
The Bun port didn't need that constraint. Code has tests.
Most enterprise domains where parallel orchestration would be valuable don't. Consider a fleet of agents reviewing contracts in parallel, each extracting key terms and checking compliance. One agent's misinterpretation of a clause propagates silently through every downstream assessment. No deterministic oracle catches it, because there is no oracle. We don't yet have much empirical documentation of what correlated failure looks like at deployment scale. We might be building the capacity to produce parallel output at extraordinary volume in exactly the territory where we have the least ability to verify it.
The test suite was the entire reason the Bun port's output could be trusted. Parallel orchestration in domains without an equivalent oracle scales the gap where verification should be.
- Princeton's reliability decomposition: The revised paper proposes 12 metrics across four dimensions showing that recent capability gains have yielded only small improvements in reliability, even for frontier models.
- Multi-agent failure patterns at scale: The MAST taxonomy, validated across 1,600+ execution traces, found that systems with explicit external verifiers had 15.6% fewer failures than those relying on peer review alone.
- The Bun port's open edges: The 99.8% figure is from canary testing, not production, and community analysis of PR #30412 has surfaced 26 known structural divergences between the Zig and Rust implementations that the test suite cannot observe.
- Correlated failure, acknowledged but unquantified: The 2026 International AI Safety Report explicitly notes that agents built on the same base model may exhibit correlated failures, while conceding that empirical evidence in deployed systems remains limited.