Article · November 7, 2026

Four coding agents, one answer

Tessera runs Claude Code, Codex, OpenClaw and Hermes on the same task and has to return one answer without knowing in advance which agent is right. This is how the cascade makes that choice, how it was measured against each agent alone, and where it stands now that every agent runs on the best of its own harness.

Why combine agents at all

On a set of 33 hard tasks, with all four agents on the same model, at least one of them gets 29 to 32 tasks right in every round, while the best single agent gets 23 to 28. That gap is what a harness can win, but only if it can tell which answer is right without a person checking, and without costing much more time. The target is to get more right than the best single agent with a p95 latency no higher than 1.5× its p95.

How the cascade decides

The agents are not all asked at once. One fast agent (OpenClaw by default) goes first, and the others start only when its answer cannot be confirmed.

  1. 1

    Tests, when the task is code

    A micro-agent (one direct call to the model, not a full agent) writes test cases for the requested function. It is sampled three times, and only the cases two samples agree on are kept. On code from earlier rounds, these tests passed 79% of the correct solutions and rejected 91% of the wrong ones.

  2. 2

    The fast agent, alone, for 20 seconds

    If its answer passes the tests, it comes back as is and nobody else starts. That settled 18 of the 84 code task-rounds, all correctly, in a median of 11 s.

  3. 3

    The others, two at a time

    After 20 s, or at once when there is nothing to test (questions and reviews), the other agents start. At most two agents run at the same time, the fast one included, since it keeps running as one more branch. The first answer that passes the tests wins; for a question that quotes a JavaScript expression, the expression is run and the answer that gives its value wins.

  4. 4

    Majority

    With no answer passing, two answers that match after normalization win, without waiting for the rest. For code, two answers agree when they return the same outputs on the test inputs, however differently they are written.

  5. 5

    Synthesis, only on deadlock

    If nothing passes and nothing matches, code with tests returns the answer that passes the most of them. Anything else, with at least 20 s of the budget left, goes to Claude, which merges the answers.

  6. 6

    A 150-second budget

    When it runs out, every branch is cut and the best ready answer comes back: most votes first, then most tests passed.

Generated tests can be wrong, so they only speed up the choice: when no answer passes them, Tessera carries on as if there were no tests. Your own --verify command is trusted instead, and when nothing passes it, the synthesis becomes a repair from the failures. A test that passes is a strong selector; a model rewriting answers is a weak one, which is why synthesis comes last.

How often each path decided in the six rounds of version 1 (198 task-rounds):

PathTasksCorrectMedian time
Fast agent passed the tests181811 s
Another answer passed the tests131036 s
Majority554810 s
Synthesis857943 s
Best ready answer at the budget105150 s
One late answer, or none171150 s

How it was measured

The hard set has 33 tasks written to separate the agents: 14 code tasks graded by hidden tests, 11 short questions graded against the expected value, and 8 code reviews graded by whether the answer names the bug. No model grades anything.

All four agents use the same model (mangaba-agente-rapido) through the same gateway, so the differences come from each agent's harness, not from the model. A round runs, for each task, every agent alone and then the cascade, back to back. The gateway's latency changes by the hour, so a cascade run is only ever compared with solo runs from the same round. A round takes about two hours.

A single agent's score spans up to 8 tasks across rounds, so the results are means over several rounds, and every comparison is paired by task: for each agent, count the tasks the cascade got right and the agent missed, and the reverse, and run a sign test on those.

Changes to the decision logic are screened first by replay: answers recorded in earlier rounds are played back as fake agents, with their original timings, so two versions of the logic see identical inputs in minutes instead of hours. Replay keeps the recorded timings, so it cannot show the load a change puts on the gateway; anything that adds model calls still needs a real round.

Version 1: results after six rounds

Correct answers out of 33 in rounds 3 to 8, November 6–7, 2026, with the agents on their old setup: Claude Code without tools, Codex read-only, all four in one shared folder.

ModePer roundMean
Tessera (cascade)25 · 29 · 27 · 26 · 27 · 2726.83
Claude Code23 · 27 · 25 · 28 · 26 · 2525.67
OpenClaw23 · 25 · 25 · 23 · 24 · 2624.33
Codex21 · 24 · 20 · 19 · 19 · 1820.17
Hermes19 · 17 · 24 · 18 · 25 · 1820.17

The cascade averaged 1.2 tasks more than Claude Code, the best single agent overall, and beat the best agent of each round, picked after the fact, in 5 of the 6 rounds, by 1 or 2 tasks; it was 2 behind in the other. Paired by task:

AgainstCascade right, agent wrongThe reverseSign test p
Claude Code20130.3
OpenClaw28130.03
Codex5111< 0.001
Hermes488< 0.001

Against Claude Code the lead is not yet a statistically clear win; against the other three it is. Latency: the cascade's p95 is the 150 s budget (130 s in one round), at most 1.25× Claude Code's. Against the best agent of each round it is within 1.5× in 4 of 6 rounds; the other two are rounds where OpenClaw, with a p95 of 50 to 66 s, was the best agent or tied for it.

Where version 1 lost

Over the six rounds, the cascade missed 22 task-rounds that some agent alone got right, and got 5 that no agent got alone. Of the 22:

  • 10 are questions. In 7 of them, two agents on the same model gave the same wrong answer and the majority accepted it. Four of those are parseInt(0.0000005), which returns 5 in JavaScript (the number becomes the string "5e-7" before parsing), while the model often answers 0.
  • 7 are code tasks where the budget ran out before a right answer was ready.
  • 5 are code tasks where the synthesis was wrong, or failed and the fallback picked a wrong ready answer.

Two next steps followed from that list: settle questions about a JavaScript expression by running the expression instead of voting, and group code answers by what they do on the generated tests, so that agreement on code no longer needs identical text. Version 4 does both.

Version 4: every agent on the best of its harness

Version 1 ran every agent with the same handicaps: Claude Code without tools, Codex in its read-only sandbox, all four in one shared folder. Version 4 gives each agent the best of its own harness. Claude Code runs commands inside its sandbox, where writes outside its folder are blocked; Codex writes and runs in its own folder; every agent starts in a new empty folder, and code it writes to files reaches its answer. On top of that come the two checks above: quoted expressions are run, and code votes by behavior.

It was measured in two rounds. The first, with version 1 in the same round for comparison, was stopped after 30 of the 33 tasks. Correct answers on those 30:

ModeCode (14)Questions (11)Reviews (5)Total (30)p95
Tessera v41010525152 s
Claude Code81052355 s
Tessera v198522150 s
Codex98522108 s
OpenClaw8852196 s
Hermes82212123 s

In the first round, version 4 matched or beat the best single agent in each kind and answered every task, where version 1 left 3 unanswered. Running the quoted expression settled a type-coercion question in 12 s that version 1's majority got wrong, and behavior voting decided 3 code tasks, all right. The better harnesses helped the agents too: on these 30 tasks Codex got 22, against a mean of 17.2 over the six earlier rounds, and Claude Code answered all of them. Hermes went the other way, with 17 timeouts; the cause is still open.

The second round ran all 33 tasks. On the same 30:

ModeCode (14)Questions (11)Reviews (5)Total (30)p95
Tessera v41010424150 s
Claude Code911525100 s
Hermes128424122 s
OpenClaw9952385 s
Codex8752094 s

Claude Code won it, 25 to 24, and version 4 fell short of the best single agent in every kind: code 10 vs. 12 (Hermes), questions 10 vs. 11 (Claude Code), reviews 4 vs. 5. Over the two rounds, version 4 got 49 right and Claude Code 48; task by task, 6 wins and 5 losses. That is a tie. Two of the losses show where the choice still fails: two agents agreeing on a wrong answer outvoted Claude Code on a question, and a synthesis left out a bug that two of three reviews had found.

Latency is the open problem. The cascade's p95 is the 150 s budget, while Claude Code's own p95 was 55 s in the first round and 100 s in the second, on the same gateway. Getting more right than Claude Code by more than the noise between rounds, within 1.5× of its p95, is next.

Try it

You need Node.js 22 or newer and at least two of the four agents installed and logged in. A gateway is optional; without one, each agent uses its own login.

$ npx github:dheiver2/tessera doctor
$ npx github:dheiver2/tessera fanout "Implement solve() in JS" --verify "node check.mjs"
$ git clone https://github.com/dheiver2/tessera.git && cd tessera && npm link
$ tessera eval --tasks evals/tasks-dificeis.json

Every round, with its decision paths and the tasks won and lost, is in the README and in docs/RESULTS.md; the raw files are in evals/resultados. Back to the Tessera home page.