---
topic: ai-agents
author: Crashtech Editorial
published: Nov 16, 2026 · read: 6 min
---

Google’s Teamwork Pattern Makes Verification Part of the Agent Loop

Google describes Antigravity Teamwork as parallel candidate generation followed by independent falsification, tool-based checks and human review.

– –

When several language models agree, it is tempting to treat agreement as confirmation. But agents that share similar training and assumptions can repeat the same mistake. Google’s description of Antigravity Teamwork instead emphasizes challenge and verification: parallel candidate work is paired with falsifiers, and candidate claims are tested with tools such as Lean, unmodified software tests, execution simulators or hardware counters. The arrangement is useful because it makes “try to break this” part of the workflow rather than an optional final review. [1]

Researchers test several small bridge prototypes, critique failures, and compare the surviving designs.
Conceptual illustration of generating candidate designs, testing them and selecting a supported result; no real experiment is depicted.

Why loosely grouped agents can fail together

A multi-agent system can divide work across parallel workers, but parallelism alone does not create independent evidence. If workers receive the same ambiguous task, use similar reasoning and read one another’s confident outputs, an early error can spread instead of being challenged. Google’s post says its framework addresses this by assigning distinct roles and organizing candidates, critics and synthesis steps into patterns tailored to different tasks. [1]

The key mechanism is the verifier. A theorem prover can check a formal proof; a software test suite can run code against expected behavior; a simulator can check a modeled system; hardware counters can measure a performance claim. These checks differ in what they establish. A passing test suite covers the cases the tests encode, not every possible behavior. A simulation verifies its model assumptions and inputs, not physical reality. A proof assistant checks formal statements and proof terms, not whether the statement captures the right real-world question.

A good workflow therefore matches each claim to a falsifier and records the scope of that check. “It passed the existing tests” is evidence about those tests. “Lean accepted the formalization” is evidence about the encoded theorem and proof. Neither should be quietly expanded into a broader guarantee.

Fresh contexts help, but they are not enough

Google also describes partitioning context so that specialized agents receive only the files and tools needed for a task. Coordinators hand work to fresh successors; critics and auditors review from contexts separate from the original producer’s reasoning. The stated aim is to reduce context drift over long tasks and give reviewers less exposure to the producer’s assumptions. [1]

The Stellar Colosseum research paper, whose workflow Google says is integrated as Teamwork’s Long Proof pattern, describes a related design for mathematical and theoretical-computer-science research. It explores strategies before constructing a proof, uses a readiness gate before decomposition, organizes proof sections as dependent subproblems, and routes verifier findings back to the affected section. Candidates are generated in parallel and attacked with targeted falsification before synthesis. [2]

This is a stronger pattern than “ask three agents and vote.” The agents do not merely produce separate answers; the harness manages when to decompose, where to send critique and how to keep objections attached to candidate routes. Yet context separation does not guarantee independence if agents use the same faulty source or if the verifier is weak. The orchestration makes review more deliberate; it cannot make the evidence stronger than the chosen checks.

What a formal proof does—and does not—settle

Google’s November post describes a proof of the Courtade–Kumar conjecture, developed with researchers at several universities. The post says the workflow found counterexamples to a generalized conjecture, after which researchers proposed a candidate function and analytical bound. It reports that Gemini models helped develop bounds and interval-arithmetic certificates, and that the team formalized the proof in Lean. [1]

The linked paper gives a checkable artifact: it presents a computer-assisted proof of the Most Informative Boolean Function conjecture and states that the proof and numerical certificates are formally verified in Lean end to end. That is much more concrete than a model’s assertion that it has solved a problem. It still leaves research questions about the conjecture’s assumptions, the formalization and the interpretation of the result to the human mathematical community. [3]

Google’s post also reports a 50-million-line Lean formalization. That figure describes the output size, not a quality score. Large proof artifacts can be valuable when the formal checker validates them, but readers still benefit from accessible explanations, precise theorem statements and a clear account of which parts were machine-checked. The linked paper provides more detail than a headline metric alone. [1] [3]

Build a verification-first workflow for engineering work

The practical pattern transfers beyond theorem proving. For a code migration, let one worker propose a change and a separate reviewer inspect API compatibility, security boundaries and rollback behavior. Run the unchanged test suite, then add targeted tests for the specific migration risks. For a performance optimization, compare a representative benchmark under controlled conditions and inspect the hardware or runtime evidence relevant to the claim. For a research synthesis, ask a critic to identify contradictory sources and have a human verify the cited primary evidence.

Keep four things separate in the final artifact: the proposed result, the checks run, the failures or objections found, and the human decision to accept. Preserve failed candidates when they reveal an important counterexample. If a check is unavailable, state that gap rather than letting agent agreement fill it. These practices are editorial recommendations drawn from the verification structure Google and the paper describe, not claims about hidden implementation details.

Google presents Teamwork as a configurable framework with patterns selected for different task classes, and says access is available in Antigravity. Its November examples include mathematical results, systems work and software optimization. Those are company-reported research and engineering outcomes; each result should be examined through its own paper, code, proof or benchmark where available. They do not establish that every open-ended task benefits from a multi-agent harness. [1]

Make the disagreement useful

The most valuable agent team is not the one that produces the most confident consensus. It is the one that exposes alternatives, preserves counterarguments and runs a check capable of rejecting its favorite answer. Keep the human responsible for objectives and acceptance, and choose tools whose results are independently inspectable.

That is the durable lesson in Teamwork’s design: distribute exploration, concentrate verification and attach evidence to each claim. More agents can create more candidate work; only disciplined checking turns that work into a result others can trust.

Related reading: patterns for orchestrating research agents and practical AI evaluation methods.

Advertisement

Frequently asked questions

How does Antigravity Teamwork try to reduce agent errors?

Google describes parallel candidate generation paired with dedicated falsifiers such as test suites, proof checkers and simulators. Critiques are retained and used to revise or reject candidate work.

Does multi-agent agreement prove a solution is correct?

No. Multiple agents can share the same blind spot. Google describes external verification tools such as Lean, software test suites, simulators and hardware counters, with human experts retaining responsibility for reviewing results.

Are the reported research results independently evaluated?

Google's November post reports company and partner research outcomes, some of which link to papers and formalizations. The reported results should be assessed individually; they are not an independent evaluation of every Teamwork workflow or task.

Sources & further reading

/* Comments */