Merget vs sequential agents: fewer repeated merge operations

A merge queue should finish the code, not maximize the number of merge operations. We replayed historical Flask branches with Merget and sequential coding agents using the same model and per-conflict budget. Branches can already contain one another’s work, so we measure that alongside direct merges, tests, time and spend.

On batch 2, Merget’s median was 354.3 seconds and $1.10 modeled spend, against 692.5 seconds and $3.28 for sequential agents. It brought in 23 of 24 branch histories with 15 direct merges and held one branch for review. On batch 1, both approaches brought in all 20 branch histories.

MeasuredNovember 1, 2026
SetupSame model and budget for both approaches
WorkloadHistorical Flask branches

Batch 2 took about half the time and one-third the modeled spend. Batch 1 included all 20 branch histories with either approach.

Batch 2 observed 49% less time and 67% lower modeled spend with Merget (57% lower at list price). It incorporated 23 branch histories with 15 direct merges and held one for review.

How to read thisMedians of 3–4 completed runs; thin lines show the observed range.

Batch 2 · flask · 24 branches6 conflicted merges in the original history, all within one group of related branches; replaying it, the agents met 7 (one the replay itself produced) and validated all 7

Direct merge operationsmerge commits madeWall timeobserved median
Sequential agents24 histories included through 24 direct mergesDirect merge operations · merge commits made: 24 directWall time · observed median: 692.5 s581.6–709.1
Merget23 histories included: 15 direct + 8 contained; 1 heldDirect merge operations · merge commits made: 15 directWall time · observed median: 354.3 s302.9–354.7
Merget without the resolver23 histories included: 15 direct + 8 contained; 1 heldDirect merge operations · merge commits made: 15 directWall time · observed median: 478.3 s410.1–484.0

Batch 1 included all 20 branch histories with either approach. Merget’s median time was 13% higher and modeled spend 19% higher (82% higher at list price). Fewer direct merges did not mean less branch history included.

How to read thisMedians of 3–4 completed runs; thin lines show the observed range.

Batch 1 · flask · 20 branches8 conflicted merges in the original history, all within two groups of related branches

Direct merge operationsmerge commits madeWall timeobserved median
Sequential agents20 histories included: 17 direct + 3 carried inDirect merge operations · merge commits made: 17 directWall time · observed median: 1,674.0 s1,512.4–1,684.6
Merget20 histories included: 14 direct + 2 contained + 4 carried inDirect merge operations · merge commits made: 14 directWall time · observed median: 1,891.0 s838.3–1,911.3

The one-call resolver cut modeled spend by 69% on the same batch outcomes. It reduced agent sessions from 7 to 2, with 26% less observed time.

26%less observed time than Merget with agents alone354.3 s versus 478.3 s
69%lower modeled spend57% lower at list price
7 → 2agent sessions with the resolver enabledsame 23 incorporated branch histories and one held branch

Flask batch 2, medians of 3–4 completed runs per configuration.

No new test failures in any completed run. Merget also preserved Git’s already-merged text in every auditable merge.

  • Clean results. No finished run left a conflict marker, an open Markdown fence, an edit outside the conflicts or a new test failure at the end.
  • Edits outside the conflict. In the audited merges, sequential agents altered git’s merged text once in each run of batch 1; Merget altered none.

A branch can arrive through its own merge or inside another branch’s history. Both count as work delivered.

Direct merge operations

A branch gets its own merge commit. A later branch can include earlier branch history without another direct merge.

Included history and validation

Contained branches enter through another branch’s history; carried-in branches arrive with a later successful merge. “Validated” counts direct conflicted merges that pass the test recipe.

One-call resolver

Merget’s first way to resolve a conflict: one model call per conflicted file, with no agent. When it is off or fails, Merget falls back to an agent session.

Modeled and list-price spend

Modeled spend charges cached tokens at the full input rate. List-price spend charges cached tokens at their discounted rate. We quote both.

Every completed run, with each unit named. Direct merges, included history, checks and resource use.

Full results: medians on every measure

Batch 1 included all 20 branch histories with either approach. Merget’s median time was 13% higher and modeled spend 19% higher (82% higher at list price). Fewer direct merges did not mean less branch history included.

How to read thisMedians of 3–4 completed runs; thin lines show the observed range.

Batch 1 · flask · 20 branches8 conflicted merges in the original history, all within two groups of related branches

Wall timeobserved medianModeled spendobserved medianList-price spendobserved medianDirect merge operationsmerge commits madeDirect conflicts validateddepends on merge orderAgent sessionsfewer is better
Sequential agents20 histories included: 17 direct + 3 carried inWall time · observed median: 1,674.0 s1,512.4–1,684.6Modeled spend · observed median: $9.74$8.66–$12.88List-price spend · observed median: $2.64$2.52–$2.98Direct merge operations · merge commits made: 17 directDirect conflicts validated · depends on merge order: 7Agent sessions · fewer is better: 10
Merget20 histories included: 14 direct + 2 contained + 4 carried inWall time · observed median: 1,891.0 s838.3–1,911.3Modeled spend · observed median: $11.56$9.61–$12.72List-price spend · observed median: $4.80$3.52–$5.34Direct merge operations · merge commits made: 14 directDirect conflicts validated · depends on merge order: 5Agent sessions · fewer is better: 16.5

Batch 2 observed 49% less time and 67% lower modeled spend with Merget (57% lower at list price). It incorporated 23 branch histories with 15 direct merges and held one for review.

How to read thisMedians of 3–4 completed runs; thin lines show the observed range.

Batch 2 · flask · 24 branches6 conflicted merges in the original history, all within one group of related branches; replaying it, the agents met 7 (one the replay itself produced) and validated all 7

Wall timeobserved medianModeled spendobserved medianList-price spendobserved medianDirect merge operationsmerge commits madeDirect conflicts validateddepends on merge orderAgent sessionsfewer is better
Sequential agents24 histories included through 24 direct mergesWall time · observed median: 692.5 s581.6–709.1Modeled spend · observed median: $3.28$2.98–$4.90List-price spend · observed median: $1.06$1.02–$1.40Direct merge operations · merge commits made: 24 directDirect conflicts validated · depends on merge order: 7Agent sessions · fewer is better: 7
Merget23 histories included: 15 direct + 8 contained; 1 heldWall time · observed median: 354.3 s302.9–354.7Modeled spend · observed median: $1.10$1.02–$1.32List-price spend · observed median: $0.46$0.43–$0.50Direct merge operations · merge commits made: 15 directDirect conflicts validated · depends on merge order: 2Agent sessions · fewer is better: 2
Merget without the resolver23 histories included: 15 direct + 8 contained; 1 heldWall time · observed median: 478.3 s410.1–484.0Modeled spend · observed median: $3.60$3.14–$3.86List-price spend · observed median: $1.06$0.96–$1.12Direct merge operations · merge commits made: 15 directDirect conflicts validated · depends on merge order: 2Agent sessions · fewer is better: 7
Full results: wall time of every run

One Merget run on batch 1 took 838.3 s, against about 1,890 s for the other three.

How to read thisEach square is one run’s wall time; runs with nearly the same time are stacked. The vertical tick marks the median.

Sequential agentsMergetMerget without the resolver
Sequential agentsBatch 1median 1,674.0 s3 runs
MergetBatch 1median 1,891.0 s4 runs
Sequential agentsBatch 2median 692.5 s3 runs
MergetBatch 2median 354.3 s4 runs
Merget without the resolverBatch 2median 478.3 s3 runs
Full results: every finished run

Every completed run: direct merges, branch history that arrived inside later merges, and checks.

BatchApproachRunWall timeDirect merge operationsArrived inside later mergesCarried inDirect conflicts validatedNew test failures
Batch 1Sequential agentsRun 11,674.0 s17 direct0370
Batch 1Sequential agentsRun 21,512.4 s17 direct0370
Batch 1Sequential agentsRun 31,684.6 s17 direct0370
Batch 1MergetRun 11,911.3 s14 direct2450
Batch 1MergetRun 1, repeated838.3 s14 direct2450
Batch 1MergetRun 21,890.2 s14 direct2450
Batch 1MergetRun 31,891.8 s14 direct2450
Batch 2Sequential agentsRun 1709.1 s24 direct0070
Batch 2Sequential agentsRun 2692.5 s24 direct0070
Batch 2Sequential agentsRun 3581.6 s24 direct0070
Batch 2MergetRun 1354.0 s15 direct8020
Batch 2MergetRun 1, repeated354.5 s15 direct8020
Batch 2MergetRun 2354.7 s15 direct8020
Batch 2MergetRun 3302.9 s15 direct8020
Batch 2Merget without the resolverRun 1484.0 s15 direct8020
Batch 2Merget without the resolverRun 1, repeated410.1 s15 direct8020
Batch 2Merget without the resolverRun 2478.3 s15 direct8020

The same runs: agent sessions, model calls and spend in both units.

BatchApproachRunAgent sessionsResolver callsAgent model callsModeled spendList-price spend
Batch 1Sequential agentsRun 110075$8.6558$2.5236
Batch 1Sequential agentsRun 210092$9.7428$2.6381
Batch 1Sequential agentsRun 310091$12.8790$2.9787
Batch 1MergetRun 11776149$11.3749$4.9979
Batch 1MergetRun 1, repeated1677134$9.6123$3.5246
Batch 1MergetRun 21776146$11.7355$4.5957
Batch 1MergetRun 31677148$12.7238$5.3383
Batch 2Sequential agentsRun 17054$4.8957$1.3978
Batch 2Sequential agentsRun 27046$2.9831$1.0561
Batch 2Sequential agentsRun 37049$3.2830$1.0243
Batch 2MergetRun 12712$1.0231$0.5048
Batch 2MergetRun 1, repeated2714$1.1477$0.4263
Batch 2MergetRun 22712$1.0489$0.4346
Batch 2MergetRun 32713$1.3223$0.4764
Batch 2Merget without the resolverRun 17052$3.8563$1.1162
Batch 2Merget without the resolverRun 1, repeated7049$3.1382$0.9619
Batch 2Merget without the resolverRun 27048$3.6001$1.0576
Full results: what one resolver call and one agent session cost

What one resolver call and one agent session cost, modeled.

$0.0177per resolver call, modeledsingle calls ranged $0.0055–$0.0995 over 345 calls
$0.89–0.92per agent session, modeled$0.8936 in a short trial run before the benchmark; $0.9193 over 161 sessions
$0.12–0.18the resolver’s cost per Merget run on batch 2 (7 calls)

Use Merget to reduce repeated merge work. The strongest comparison isolates the resolver on the same branch outcomes.

The one-call resolver handled enough of batch 2 to replace five of seven full agent sessions. With the same included branch histories, median modeled spend fell from $3.60 to $1.10 and time fell from 478.3 to 354.3 seconds.

Against sequential agents, branch containment avoided repeated direct merges: on batch 2 Merget brought in 23 branch histories with 15 merges.

Next step

Cut the repeated merge work. Connect one repository and let Merget take the rebases your agents repeat.

Sign up free
How we ran it

Measured November 1, 2026 on historical Flask branches: the same model and per-conflict budget for both approaches, 3 to 4 runs each.