Ask four agents.
Keep the answer that works.

Tessera is a harness for coding agents. It runs Claude Code, Codex, OpenClaw and Hermes on the same task, each on the best of its own harness, and keeps the answer that passes the tests; with no tests, the one most of them agree on. Open source, MIT license.

$ npx github:dheiver2/tessera fanout "Implement solve() in JS" --verify "node check.mjs"
[done] openclaw 10912ms verification: passed
[cancelled] claude 11020ms — dismissed (an answer was already chosen)
[cancelled] codex 11024ms — dismissed (an answer was already chosen)
[cancelled] hermes 11031ms — dismissed (an answer was already chosen)
branches: 11031ms gateway: each agent's own
=== final answer (openclaw passed verification, no synthesis, 11040ms total) ===

OpenClaw passed the tests in 10.9 s and the other three were cancelled. Works with any two or more of the four agents.

How Tessera decides

Proof first, then consensus

The fast agent goes first; the others start only when its answer cannot be confirmed. Tessera decides in the order below and cancels whoever is left behind.

1Verification

The answer that passes the tests

Each answer is tested as soon as it arrives, with your --verify or with tests a micro-agent writes. The first one that passes ends the fanout and comes back untouched. A question that quotes a JavaScript expression is checked by running it.

2Majority

Two matching answers win

When nothing passes, Tessera compares the answers: the same text for questions, the same outputs on the test inputs for code. When two match, that one comes back right away, without waiting for the other agents.

3Synthesis

Only on deadlock

If nobody passed and nobody agrees, code with tests returns the answer that passes the most of them; anything else is merged by one agent. If the synthesis fails, a finished answer comes back instead of nothing.

Commands

Works with the agents you already use

Plain Node, no dependencies, no install. Any two or more agents; each uses its own login, or all go through any gateway you choose.

$ npx github:dheiver2/tessera doctor # finds your agents, Node 22+
$ npx github:dheiver2/tessera fanout "your question"

tessera sync

One configuration (rules, skills, MCP) written in each agent's format, without touching what is yours.

tessera fanout

The same task to all four, with verification, majority, and synthesis only on deadlock.

tessera pipeline

Chain agents: one plans, another implements, another reviews.

tessera do · delegate

Send a task to a single agent, in the foreground or in the background.

tessera eval

Measure each agent alone and Tessera on the same tasks, with private tests.

tessera doctor

Shows what is installed, what is in sync and what is missing.

Measured results

Two rounds: a tie with Claude Code, 49 to 48

Every agent on the best of its own harness (Claude Code with its terminal in a sandbox, Codex writing in its own folder), all on the same model, all modes in the same round. On the same 30 tasks, the cascade won the first round, 25 to 23, and Claude Code the second, 25 to 24. Latency is the open problem: the cascade's p95 is the 150 s budget.

ModeRound 1Round 2Both (of 60)p95, rounds 1 · 2
Tessera v4 (cascade)252449152 · 150 s
Claude Code23254855 · 100 s
OpenClaw21234496 · 85 s
Codex222042108 · 94 s
Hermes122436123 · 122 s
49 · 48

correct of 60 for the cascade and for Claude Code, the best single agent: the same 30 tasks in two rounds, a tie

6 · 5

tasks won and lost against Claude Code, task by task; the cascade answered 59 of 60

1.5–2.8×

p95 against Claude Code with its tools (the 150 s budget vs. 100 s and 55 s): latency is the open problem

Before: v1 over 6 rounds (Nov 6–7, 2026, 33 tasks)

With the agents on their old setup (Claude Code without tools, Codex read-only, one shared folder), the cascade was 1 or 2 tasks ahead of the round's best agent in 5 of 6 rounds; against Claude Code, task by task, 20 wins and 13 losses (p = 0.3).

ModeCorrect per roundMean (of 33)
Tessera v1 (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

How each answer is chosen and how it was measured: the article. Raw results, tasks and the aggregation script are in the README. Reproduce with tessera eval.

Try it on your own tasks

Run tessera eval on your tasks and see whether, in your case, four agents beat one.