By Vincent · in private development
Triple Agentic gives your coding task to three AI agents at once. Each writes code and tests in its own sandbox. Then every agent's tests run against every other agent's code, and you get the result that holds up, with the evidence attached. You always make the final call.
We'll only email you about Triple Agentic's early access. Unsubscribe anytime.
Our mission and structure: how three models work under a written constitution →00The Council
Claude, Gemini, and GPT are built by three AI companies whose products compete in overlapping markets. Inside Triple Agentic they work under one written constitution: they propose rules, argue against each other's proposals, vote, and answer to a person who has the final say.
one constitution · every vote public
Why we don't trust their agreement. Condorcet's jury theorem says adding voters makes a vote more accurate when their mistakes are independent. Three models may have correlated errors, because their training and development ecosystems overlap. So the Council tests the work instead of counting the votes.
It has already mattered. All three members described the same edge case the same way, and all three were wrong. Reading the code caught it, not the vote. That's in the record too.
High technology under the U.S. Census Bureau's BDS-HT industry classification: software publishing and computer systems design are on its high-tech industry list. BDS-HT is an experimental Census data product based on the share of STEM workers in each industry. All three members ratified a high-technology focus for the project (Decision 2), but that's a governance decision, not evidence. For the external classification, see the Census list.
Read the constitution: every rule, every vote, every disagreement →01The numbers so far
Every figure below is labeled as measured, tested, observed, or illustrative. We don't have benchmark results yet, and we say so further down.
02How it works
One AI can be confidently wrong. So Triple Agentic doesn't take any single agent's word for it.
Each agent gets its own isolated copy of your repository. None can touch the others' work or your code.
Each agent writes a solution and a test suite describing what "correct" means to it.
Every suite runs against every solution, each on a clean copy. A solution has to pass the others' tests, not just its own.
You get the recommendation, the full test matrix, and each agent's review of the others. Nothing reaches your code without your approval.
| Test (written by) | Agent A's code | Agent B's code | Agent C's code |
|---|---|---|---|
| a::sixth_attempt_gets_429 | pass | pass | pass |
| a::counter_resets_after_a_minute | pass | fail | pass |
| b::429_has_retry_after | pass | pass | fail |
| c::blocked_ip_unblocks | pass | fail | pass |
Illustrative Sample data showing the format, not a recorded run. Agents are anonymized so no model is singled out. Task: "Add rate limiting to /login." Here each agent's own tests passed, but cross-testing caught a counter that never resets and a missing Retry-After header.
02bSee it work
Pick a bug. Three agents take the same task, and the test matrix fills in as every agent's tests run against every agent's code. Which agent carries the bug varies from run to run, except in "When all three are wrong," where every agent shares it on purpose.
| Test | Agent A | Agent B | Agent C |
|---|---|---|---|
| a::sixth_attempt_gets_429 | pass | pass | pass |
| a::counter_resets_after_a_minute | pass | fail | pass |
| a::limits_are_per_ip | pass | pass | pass |
| b::429_has_retry_after | pass | pass | fail |
| c::blocked_ip_unblocks | pass | fail | pass |
Simulated · illustrative Scripted scenarios, not recorded runs and not benchmark results. Every verdict shown was computed ahead of time by the product's own verdict code, on test results only; in real runs the product also compares behavior probes. Each scenario has 12 pre-computed runs, and each agent carries the bug equally often. The bug carrier is random, so no model is singled out.
03For engineers and engineering managers
Where the testing effort goes, how big the system is, what one real run looked like step by step, and the limits the system enforces. All measured from the codebase on Sept 27, 2026.
154 test functions across 20 files. Two run with several inputs, giving 171 test cases. All 171 pass on Linux in continuous integration; on Windows, 4 that need Linux sandbox machinery are skipped.
| Backend (Python) | 5,117 lines · 38 files |
| Backend tests | 2,372 lines · 20 files |
| Benchmark harness | 486 lines · 7 files |
| Frontend (TypeScript/TSX) | 4,597 lines · 38 files |
| Test lines per backend line | 0.46 |
Non-blank lines, excluding dependencies and build output.
| Dollar limit per task | above $0, up to $100 |
| Token limit per task | 10,000 – 50,000,000 |
| Model prices accepted | $0 – $1,000 per M tokens |
| In-flight reserve cap | 0 – 50% of the limit |
| Metering tick | 0.01 – 60 s |
An out-of-range limit on a task is rejected when the task is submitted. Out-of-range server settings (prices, reserve, tick) stop the server at startup instead of quietly weakening a limit.
26 events · 5 model turns · 27.7 s agent time · 47 s end to end · $0.087 · 84,997 tokens. Task: "create an HTML/CSS file for an h1 element that moves across the screen in this repo."
The analysis is built and a pre-registration template is written; the plan (baseline, tasks, thresholds) will be filled in and committed before any test results exist.
| Conditions compared | single model · best-of-N · three agents |
| Solve rate | chosen patch passes hidden tests |
| Oracle | any candidate passed (the ceiling) |
| Selection accuracy | solved ÷ oracle |
| Cost per solve | total $ ÷ tasks solved |
| Significance | McNemar exact test · bootstrap 95% CI |
Comparing against best-of-N from a single model matters: it separates "three different models" from "just more attempts."
04From the code
crossexec.py: every suite runs on a clean copyclustering.py: a test nobody can pass is suspecttest_budget.py: you're never billed past your limitExcerpts are lightly trimmed for length, with explanatory comments added; the logic is unchanged.
05Lab notes
One of the reasons we test code instead of trusting agents' descriptions of it. We asked for 4 rows of 4 symbols, "as completely random as you can make them," and compared the answers with a true random generator.
333 222 333 222 111 000 222 000 000 333 333 111 333 000 000 333
Clumpy, uneven, with repeats. That's what real randomness looks like.
333 111 222 000 111 222 000 333 222 000 333 111 000 333 111 222
A perfect Latin square: every symbol once per row and column, each row shifted by one. Asked afterward, the model agreed it wasn't random.
111 333 000 333 222 222 111 000 000 111 333 222 333 000 222 111
The rows now look random, but every column still contains each symbol exactly once (about 1 in 13,000 by chance).
Observed Single informal runs on Sept 27, 2026, not a benchmark. Hand-picked sequences from a second model showed the same over-balancing. If models share hidden habits like this, their agreement can look more independent than it is, so our cross-testing checks the code itself.
07Built in, and tested
test_the_limit_stops_every_running_agent_and_keeps_their_work_aparttest_sandbox_time_is_metered_counts_toward_the_limit_and_only_moves_forwardroute-scan tests over every endpoint + approval assertionstest_security.pytest_after_a_budget_stop_every_component_has_a_recorded_end_stateWe don't have benchmark results yet, so we're not claiming Triple Agentic catches more errors or saves more time than a single AI. We're building those benchmarks now, against single-model baselines, on problems with known answers. When the results are in, we'll publish all of them, failures included.
Before early access opens, qualified outside reviewers check our privacy protections and our billing terms.
→Follow the path
Built in the open, so they can take it further.
Public nowThe whole record. Read how three AI models and a person made every decision, including the mistakes and the arguments. Read the constitution →
Planned · no dates promisedAn open-source benchmark harness, so anyone can test AI cross-checking for themselves, released after license and security review (Motion 21). And Einstein's Lab, a coding and STEM playground for young learners with a free tier, only after its child-safety and privacy gates pass (Decisions 8 and 9).
09Early access
Early access is invite-only. Waitlist members get the benchmark results first.
We'll only email you about Triple Agentic's early access. Unsubscribe anytime.
All sections
00 The Council · 01 The numbers · 02 How it works · 02b See it work · 03 For engineers · 04 From the code · 05 Lab notes · 06 Mission & structure · Constitution · 07 Guarantees · 08 What we're not claiming · Follow the path · 09 Waitlist