The listing's claim is that agents have moved the bottleneck: code is cheap, and testing and verification are now the constraint. Its three symptoms are more and larger PRs eating review time, agent-written tests that cover the obvious path, and subtle bugs that survive both. The answer it promises is "a fundamentally different approach", which from Antithesis means deterministic simulation testing: run the whole system in a deterministic hypervisor, inject faults, explore schedules, and replay any failure exactly.

This site has argued the same premise from the other end. If an agent writes the code, the artifact that carries intent is the spec and the gates, not the code (The Oracle Is the Deliverable). A test an agent writes against its own implementation shares its blind spots; an oracle that has never been seen to fail is not yet an oracle (Four Controls for an Agent Skill). Deterministic simulation is one way to get an oracle the agent did not write: the properties are stated once, and the search for violations is done by the platform.

The AWS half is AI-DLC, AWS's name for an AI-driven development life cycle: AI plans and asks for clarification, humans validate in "mob" sessions, and work runs in short "bolts" across Inception, Construction and Operations. AWS pitched the same program at Agents of Impact in September.

• Who writes the properties? If the agent writes both the code and the assertions Antithesis checks, the oracle has the same author as the bug. What does Antithesis recommend for keeping properties human-owned?
• A found bug arrives as a replayable trace. Can an agent consume that trace and propose the fix inside the same loop, and how is the fix then re-verified on the schedule that broke it?
• AI-DLC puts humans in "mob elaboration" at Inception. Is that where the properties get written, or are they left to Construction?
• For a browser front end, Bombadil is Antithesis's open property-based tester. How does it relate to the hypervisor product: same property language, or separate tools?
• Bombadil has no select-option action and records no network traffic; a spec has to add both itself, which the page-model run below did. Are either on the roadmap, and should one bad extractor value or action fail a step rather than end the run?
• Bombadil's hash_current is a SimHash of JS edges. Could a page model's coverage (controls, operations, transitions) guide exploration alongside it?
• Would running Bombadil inside the deterministic hypervisor make timing findings, like states caught mid-navigation, replay exactly?
• Observability is listed as a theme. A simulated run has a perfect record of time; production does not. What carries over?

The experiment behind the questions below runs OpenAPI codegen in reverse. One Bombadil exploration records every visible control per state, including the role and accessible name Playwright would locate it by, its form or repeated item, and the requests the page sent. A miner joins form submits to the OpenAPI request schema by value and generates Playwright page objects. Their field names and types come from the body key each field was seen to fill. A Playwright test written only against the generated classes passed. The same join also produced a drift report:

• fields the schema defines but the UI cannot set;
• request values the client hard-codes;
• six of seven selects with no accessible name, reachable only by position.

Bombadil 0.7.10 ended the run twice instead of failing the step:

• an extractor returned half a surrogate pair;
• a generated click had a negative coordinate.

The same model could carry several more tests. The aim is the same as codegen's: a UI or API change is a regenerate and a reviewed diff, not hand-fixed selectors.

Server-rendered apps need one change to the recorder. A form that submits by navigation never goes through fetch, so the spec would record the FormData at the submit event instead. The form's action and method give the endpoint without waiting for a request.

Agents and oracles:

• The Oracle Is the Deliverable — the premise of this evening, argued for Q3 2026: specs, formal methods, adversarial review and hermeticity as the deliverable when agents write the code.
• Four Controls for an Agent Skill — talk proposal: an agent skill as a unit under test, with a bounded scope, an oracle seen to fail, a sealed spec, and a held environment.
• Elenctic Vibe Code Review — reviewing agent-written code by structured questioning, the human side of the "too many large PRs" symptom.
• Shan Lu at PAgE 2026: can coding agents write verifiably correct software? — agents emitting Verus proofs against human-written specs; the spec is the trusted base.

Antithesis tools, used here:

• Goldberry: Frontend Invariant Catalog for DST Harnesses — a taxonomy of UI invariants written for deterministic simulation testing with Antithesis.
• Bombadil for SPAs — running notes on Antithesis's open-source property-based tester for web apps.
• Isolation Makes the Error Channel an Oracle — serving each tool from a bare harness so the browser's error channel can act as an oracle; checked against Bombadil's published defaults.

Property-based and formal methods:

• Property-Based Test Shrink Discipline — why a property failure on one seed is signal, and pinning the shrunk counterexample.
• Property-testing the crowsnest client — a model-based oracle and drift sentinel that caught a real wire-contract change.
• The Four-Boundary Spec Mapping — where contracts, session types and formal specs each sit.
• TLA+ for System Design — the design-time half of the same argument.

• BugBash 2026 — Antithesis's own correctness conference (Washington DC, April); recordings include Will Wilson's "We won, what now?".
• Software Should Work 2026 — independent reliability conference with deterministic simulation and formal methods on the program.
• P99 CONF 2026 — this week; TigerBeetle, a deterministic-simulation shop, has a keynote, and Closed-Loop Performance Engineering covers agents doing performance work.
• Agents of Impact 2026 — AWS's Boston pitch for AI-DLC and AgentCore, a month earlier.
• AGNTCON + MCPCon North America 2026 — evals and tamper-evident audit trails for agents.

Notes pending.