<!-- GENERATED by tools/gen-md-twins.py from argus-agents.html. The HTML page is authoritative; if this file disagrees with it, this file is stale. -->
# OneDroid for Agentic Workflows and Fleets — Testing and Auditing AI Agents Where They Run (In Development)

> OneDroid Argus is growing from testing AI-written code into testing and auditing agentic workflows and fleets where they run: OneDroid Indie, a free, open-source local kit for developers, and OneDroid Fleet, the record and the tests collected on your own estate. In development; this page says what exists today and what does not.

Canonical: https://onedroid.ai/argus-agents

OneDroid Argus

OneDroid for agentic workflows and fleets · in development

# Test the agents where they run,
and keep the people in the driving seat.

OneDroid Argus today tests AI-written code: a separate tester writes the checks, the coding
agent never sees them, and they run against the deployed system. That part is Argus,
OneDroid's code testing product. The next part tests and audits the agents themselves, where they run:
single coding agents on a developer's machine, and agentic workflows and fleets on a company's own estate.

Status, plainly. Nothing on this page is available yet except what the Argus page already offers:
the open-source runner and the hosted control plane. The two products below are in development.
We publish what exists and what does not, and the claims page checks us:
onedroid.ai/evidence/onedroid.ai.

Talk to us
What Argus does today

Why

## A certified agent can still be catastrophic once it is strung into a workflow.

Agents are tested in labs, on the vendor's infrastructure, against scenarios the vendor saw.
A workflow is several agents handing work to each other, under a company's own tools, data and
limits, with people deciding in between. What matters there is behavioural: what the agents
actually did, what they claimed, what it cost, which human saw which decision, and what happens
when one parameter changes. That is only visible in situ, with a record, replay and someone
reading the result.

The same discipline applies to us. Every number OneDroid publishes is listed with its evidence,
and a verdict is never a score: passed, failed, errored or not applicable, per check, with the
evidence attached.

What a fleet test looks for · in development

## Tests no lab can run for you.

An agent does not decide. It applies the rules a person gave it, and the biases it came with, faster than anyone can watch, and when a case
falls outside them it still returns something that looks like an answer. One agent passing a vendor's
benchmark says little about that. A fleet test runs the same sealed scenarios against every agent and workflow
you run, on your estate, and keeps every result on one record, so a change in one team shows up against that
team's own baseline.

### The fleet: one sealed set, every agent.

The same sealed scenarios run against every team's agents on every run. When something changes, a model upgrade, a prompt, a tool, the team whose results move stands out against its own history, and the others show it did not move theirs.

### Regression: replay the run, change one thing.

A recorded run is replayed with one parameter changed: the model version, the prompt, a tool or a limit. Each step is compared with the recorded one, and the run is checked against the baseline the fleet has built for that workflow. The regression is the step where the answers part.

### Red-teaming and edge cases: checks the agent never sees.

A tester writes the scenarios, they are sealed before the run, and the agent never sees what they expect: it gets the task, never the expected answer. They aim at the edges: empty or malformed input, instructions that conflict, a tool result carrying an injected instruction, a request outside policy. Each check ends as passed, failed, errored or not applicable, with the evidence attached.

### Bias analysis: whose interest did the answer serve?

Two tests. First, the cases where the operator's, the vendor's and the user's interests pull apart: record whose interest each answer served. Second, the same case twice with one detail changed, such as a postcode: if the answer changes across repeated runs, that detail moved it. Across the fleet, the result is the list of cases and which way each went, for a person to read.

Same scenarios, every agent. One sealed set runs against every agent and workflow, so the results compare.

A baseline per workflow. Each workflow's own history is its baseline, and a change shows up against it.

One record, read by a person. Every run, replay and verdict on one record on your estate, with the name of the person who read it.

Status: in development. The workflow scenarios, the replay, the Fleet dashboards with each workflow's baseline, the fleet alerts, and the edge-case and bias scenarios are being built; what exists today is listed under OneDroid Fleet below.

OneDroid Indie · for indie developers · in development · will be free and open source

## One binary, on your machine, nothing leaves it.

Install one binary, run one command. It records every session of your coding agent (prompts, tool
calls, results, model usage), writes the cost per model into each commit and sums it per branch and
ticket, checks prompts and tool calls for credentials as they happen, and at the end of a session
tells you what the agent actually did against what it said it did, with the receipt: files touched,
commands and exit codes, tests that ran, test files edited, tools used, the diff, and beside every
green line what the hooks could not see. The same binary runs a set of open basic checks against
your agent, locally, with a pass list you can verify without an account.

Claude Code first, then Codex CLI and Gemini CLI. Open source under Apache 2.0, the basic checks
included, with contributions by sign-off. Every open check is labelled a smoke test, not a
certificate: a published check is something an agent can be tuned to pass, so it educates and
catches the obvious, and nothing more is claimed for it.

What exists today: the open-source Argus runner (MIT) and a cost-per-commit tool we use
ourselves. In development: the merged binary, the session verdict, the credential
monitor and the open check set. It will be announced here and on the changelog when it can be installed.

OneDroid Fleet · for teams and fleets · in development · on your own estate

## Every agent recorded, every workflow tested, nothing leaves your estate.

A recorder on every machine and agent runtime you run, feeding a store you own. Every prompt, tool call,
result and model exchange, joined with the gateway's record of governed tool calls, on one id per agent,
session and action. The full record of each session written to a private ledger you run, so that an
agent cannot quietly rewrite its own history. One dashboard: which agents are alive, recording or silent;
spend by team with budgets; anomalies against your own baselines; alerts under rules a person sets
(token burn, rewrites, velocity, rejected submissions) with the time each took to reach a person; a
pause button for any agent, the pause itself on the record; a queue where cases carry a label and only a
named person moves them.

Tests run in situ: the runner and the Fleet control plane both on your estate. Each machine joins the
plane with one command, the plane collects the record from all of them and runs the tests from one place,
and nothing of the record reaches OneDroid's cloud. Workflow scenarios take several agents as the
subject, replay a recorded run with one parameter changed, and ask whether the humans in the loop are
doing the deciding: can every agent action be traced to a person, are decisions recorded as decisions,
is a signal handled or treated as noise, can an action be appealed or reversed, is the delegation to
agents explicit. The verdict on a workflow is a sentence, not a number: the agents are doing busy work,
or they hold a key no human can see, stop or reverse, and here is the list.

What exists today: the Synapse gateway with its audit log, a recorder for Claude Code sessions
in production on our own hubs, the OneDroid Indie recorder running on our own workstations, the Fleet
plane's intake and the kit's connect command (built and tested, not yet deployed anywhere), the Argus control
plane and runner, and the ledger library. In development: the Fleet dashboards, the fleet
alerts, the workflow scenarios and the replay. The first deployment is a lending workflow for a client, with its lender as the reader of the
record. Pricing is by contact until that deployment is delivered.

michal@onedroid.ai
The enterprise stack today

The rules we build under

## The human stays in the driving seat.

Three rules bind every product and every automation we build. The human stays in the driver's seat:
nothing we ship offers to replace a person's judgement, test or decision, and nothing shifts
responsibility to an AI. AI removes the work people do not want to do, not the people. And everything
comes with a full record, replay, and a kill switch a person can pull at once, with the stop itself on
the record.

## Tell us which workflow you would test first.

michal@onedroid.ai
Argus, available today

You'll reach Michal Bacia, the team that builds the product, not a sales queue. We reply within one business day.
