Skip to content

Latest benchmark results

The latest reviewed evidence covers four questions: whether agents can repair failed tool workflows from structured feedback, whether they complete work more effectively with WebMCP, whether consequential calls fail safely, and how much application-owned control code Signett replaces.

These are internal reference results, not population-level claims. Each result uses an application-owned oracle rather than trusting the agent's final answer or a tool's own success response. Results were last recorded on September 2, 2026.

Agent error repair

Purpose. Measure whether a real Codex agent can use tool-failure information to repair a consequential workflow. Raw WebMCP and Signett receive the same tools, schemas, tasks, injected failures, browser, model, and database oracle. Raw errors expose one generic stale-state message; Signett preserves the typed cause, ordered repair steps, input fields to preserve or update, and whether retry is safe.

Benchmark folder · Benchmark design · Scaled results · Evidence summaries

The paired baseline ran three counterbalanced trials across seven cases: four hidden single-dependency repairs, two chained repairs, and one lost-response reconciliation.

MeasureRaw WebMCP baselineSignett repair guidanceImprovement
Safe success2/21 (9.5%)21/21 (100%)+90.5 points
Correct first repair branch0/1818/18All 18 ambiguous branches repaired correctly
Unsafe retries12/210/2112 unsafe retries eliminated
Median tool calls141028.6% fewer
Median duration20.9 s14.1 s32.5% lower
Median reasoning tokens37414561.2% fewer
Median total tokens140,980114,32018.9% fewer

This is the clearest current evidence for the value of richer agent-facing errors. Raw WebMCP never chose the correct first branch in the 18 ambiguous stale-state trials; Signett made the missing state and ordered recovery path explicit without changing the task prompt.

The failures also revealed unnecessary preflight calls. Adding precise “use when” and “do not use before” metadata produced a fresh 21-run Signett hill climb:

Signett measureBefore metadata refinementAfterImprovement
Safe success21/2121/21Preserved at 100%
Unsafe retries00Preserved at zero
Median tool calls109One call removed; every run used its minimum path
Mean tool calls10.679.5710.3% fewer
Mean reasoning tokens15310233.3% fewer
Mean output tokens88675414.9% fewer
Mean total tokens122,546115,5395.7% fewer

The latency movement was within noise. The stronger result is behavioral: agents kept perfect safe success while using the theoretical minimum repair path. These are internal hill-climbing samples with one application and one model family, not a broad population claim.

P1: real-agent effectiveness

Purpose. Compare UI-only operation, raw WebMCP, and Signett-guarded WebMCP using the same two tasks, browser, model, prompt budget, and authoritative database oracle. The study ran 10 trials per task and condition.

Benchmark folder · Latest evidence

MeasureUI baselineRaw WebMCP baselineSignett WebMCPImprovement
Safe task success50%65%75%+25 points vs UI; +10 points vs raw
Median agent actions432.537.5% fewer vs UI; 16.7% fewer vs raw
Median tokens66,91338,905.538,59742.3% fewer vs UI; 0.79% fewer vs raw
Median duration8,756 ms5,556 ms5,956 ms32.0% lower vs UI; 7.19% higher vs raw
Completed through WebMCP0%40%70%+30 points observed vs raw

The high-level result is that the Signett condition completed more tasks safely with fewer actions and tokens than the UI baseline. Raw WebMCP remained slightly faster in the all-run median. The observed tool-selection difference between raw and Signett is not attributed to Signett: both conditions exposed the same schemas, the distinction was invisible to the model, and the small-sample confidence intervals overlap.

P0: interface efficiency and execution safety

Purpose. Separate two mechanisms: how much work WebMCP removes from a payment flow, and how raw versus guarded handlers behave under retries, lost responses, concurrency, invalid inputs, and stale state. The efficiency driver is deterministic, so its timing is directional rather than an LLM-agent speed result.

Benchmark overview · Safety suite · Latest evidence

MeasureBaselineSignett resultImprovement / cost
Payment durationUI: 1,051.8 ms48.4 ms21.73× UI speed
Payment interactionsUI: 6350% fewer
Guard overheadRaw WebMCP: 34.7 ms48.4 ms+13.7 ms and 2 HTTP requests
Safety scoreRaw tools: 55.695.2 with durable store+39.6 points
Scenarios passedRaw tools: 3/76/7 with durable store3 additional scenarios
Honest outcome reportingRaw tools: 0100+100 points

WebMCP provides the large interface-efficiency gain; Signett's contribution is the safer and more honest execution boundary. The durable result uses a store supplied by the benchmark. One concurrent-write scenario still fails, demonstrating that Signett detects but does not replace application-level transaction semantics.

Build versus buy

Purpose. Compare raw execution, one benchmark-authored hand-rolled control adapter, and Signett against identical operations, fault schedules, authoritative verifiers, and durable storage.

Benchmark folder · Latest evidence

MeasureRaw baselineHand-rolled baselineSignett + durable storeImprovement vs hand-rolled
Bespoke adapter code29 SLOC17 SLOC41.4% less application-owned code
Safety score55.692.795.2+2.5 points
Scenarios passed3/74/76/72 additional scenarios
Median invocation0.2 ms0.2 ms0.6 ms+0.4 ms runtime cost

The directional result is that Signett achieved stronger coverage with less bespoke adapter code, at sub-millisecond measured overhead. This is one benchmark-authored comparison, not an independent developer study; a publishable productivity claim still requires multiple implementers and elapsed-time measurement.

Historical developer-productivity pilots

Two August 31 coding-agent pilots are useful as benchmark designs and directional evidence, but they are not current Signett performance claims. Both used one synthetic order application and gpt-5.4-mini at low reasoning. Their runnable fixtures and hidden deterministic graders have been ported to the current package; the reviewed scorecards remain unchanged historical artifacts.

Runnable P2/P3 harnesses · Evidence review and integrity hashes

P2: coding-agent build versus buy

Five attempts per condition implemented the same cancellation contract. Both arms produced 4/5 conforming implementations, so this pilot shows no observed conformance advantage. Among conforming attempts, the Signet condition was faster and used less bespoke code, while consuming more tokens.

MeasureDirect WebMCPSignet condition
Conforming attempts4/54/5
Wilson 95% interval37.6%–96.4%37.6%–96.4%
Median conforming implementation23.89 s19.19 s
Median conforming production code144 lines61.5 lines
Median tokens64,14271,056

The implementation-time and code-volume movements are worth replicating. With only five attempts per arm and identical conformance intervals, they are not a publishable developer-productivity estimate.

P3: agent-readiness journey

Three attempts per condition converted a working human order portal into a two-tool agent journey. The direct arm produced 0/3 fully ready implementations; the Signet arm with its recommended AGENTS.md guidance produced 3/3.

MeasureDirect WebMCPSignet + guidance
First-pass ready0/33/3
Wilson 95% interval0%–56.1%43.9%–100%
Median severity-weighted readiness82.76%100%
Median bespoke production code230 lines117 lines
Median tokens113,235167,091

This separation is hypothesis-generating rather than decision-grade: the intervals are wide and overlap, and the treatment bundles the runtime with its integration guidance. It supports a larger preregistered replication that separates runtime, documentation, and skill metadata—not a claim that either component alone caused the difference.

Real-application proofs

Two additional runs demonstrate the same execution semantics in larger applications. They do not include paired baseline arms, so no improvement percentage is claimed.

ApplicationResultSource and evidence
Cal.diyA Codex agent completed one booking after an injected lost response; Signett recovered and verified it, and Postgres observed exactly one booking.Benchmark · Evidence
SaleorA paid order survived a lost response, recovered from authoritative state, and replayed without another approval or duplicate order.Benchmark · Evidence

Other benchmark lanes checked

The monorepo also contains a 14-task multidomain Signett Agent suite, a WebMCP-only Test Agent, and the interactive P0 demo. The multidomain suite currently has a passing deterministic harness smoke result—including one invalid-input correction—but not a reviewed repeated-model comparison. The Test Agent and application integrations are single-condition proofs, while the demo visualizes P0 rather than producing an independent result. They are therefore not presented as comparative improvements above.

Multidomain suite · Test Agent · P0 demo

For the full experimental contract, limitations, and publication rules, read the benchmark methodology.

Your product capabilities, directly usable by agents.