One step, inserted before the action.Give agents a model to reason with.
Agents can act, but they can't test the consequences first. Ergodic builds a running model of your operation and hands it to the agent as calls it can make before it commits: what is true now, why it moved, and what would happen if it did this.
Reroute the Tuesday run via Riverside
2,000 runs against your operation as it stands. p10 to p90 on service: +22 to +36 points.
One call, made before the agent acts. The agent keeps the decision, it just stops making it blind.
The world is built from whatever you already run.
Ergodic plugs into the systems that already hold your operation, turns what they describe into a working model of it, and keeps that model current as things change. If it emits a record, it can feed the model.
Systems of record
ERP, WMS, MES, CRM and the warehouse, through their own interfaces. We read what they already write, so nothing needs re-platforming before you start.
Timed from the start: what happened, and when you knew about it.
Spreadsheets, and the work around the system
The allocation rules, the supplier scorecard, the planner's workbook. Operations run on these as much as on the ERP, and they carry the constraints nobody wrote into software.
They feed the same picture as any other source.
Devices, telemetry and documents
Line sensors, vehicle and handheld telemetry, service tickets, contracts and email threads. This is usually where the evidence for the real constraint lives.
Whatever emits a record can feed the world an agent reasons over.
Three things an agent can ask before it decides.
There is no separate agent product. Atlas, Root Cause Analysis and Equinox are one framework with three parts, and an agent reaches them the way a person reaches the console. A query costs a call, not a turn of the agent's own reasoning, so it can ask before it acts.
What is true right nowAtlas
What you run and how it is doing right now: this supplier's real lead time, this warehouse's cover once in-flight orders settle, which customers sit downstream of a route.
Straight answers with their evidence, not prose the agent has to unpick.
Why it movedRCA
Instead of reasoning through raw logs, the agent asks for a verdict: the driver, the chain to the symptom, the evidence behind it. Where the evidence is thin, the answer says so.
That refusal is what stops an agent inventing a cause to fill the gap.
What happens if I do thisEquinox
The agent proposes an action and gets it simulated against your operation as it stands: the outcome spread, the measures it moves, and whether your constraints permit it at all.
Infeasible actions come back blocked, with the rule that blocked them.
As tools and skills
The three calls arrive as tools and skills in whatever framework your agent already runs on. It decides when to ask, asks mid-task, and asks again once the answer changes what it was going to do. This is the way to run it.
As data in the prompt
Where an agent can't take new tools, the framework runs first and its answers are written into the prompt the agent already reads. You lose the mid-task question, and the agent still acts on what the model knows. It is the route some of our benchmark runs were required to use.
Either way the agent has more to go on than it does alone. What that changed on each benchmark, including where it changed little, is set out with the method notes. See the results →
Where the step goes in the loop.
An agent loop is look, decide, act, look again. Left alone it acts first and finds out afterwards. The simulate step goes between deciding and acting, and the real outcome is written back, so the next loop starts better informed.
Your agent still looks, decides and acts. Ergodic adds the simulate step in the middle, and writes the real outcome back for the next loop.
Safety isn't a review step. It runs continuously.
Two things have to be true before an agent is trusted with consequential work: problems have to surface before they land, and the rare, ugly scenario has to have been tried somewhere other than production. The platform does both, on the same model the agent reasons over.
Problems found proactively, improvements proposed automatically
The model runs your operation forward continuously, so a breach shows up while it's still weeks away. Each one arrives with a proposed fix that has already been simulated, so the choice is to apply it, adjust it, or hold it for a person. What you approve becomes evidence the next proposal is judged against.
Safety scenarios run against the model, not the business
Rehearse the events you can't afford to meet unprepared: a supplier lost, a plant down, a demand spike, a rule change. Each is run against the model rather than the business, and the ones that breach your tolerance come back with the response that held and the cost of holding it. That set becomes a standing check, re-run as the operation changes.
Actions your business can't take come back blocked
Constraints are part of the model, not a policy document beside it. An action needing stock you don't have, a supplier you haven't approved, or cash beyond the limit is refused with the rule that refused it, before the agent can commit to it.
The same check applies whether a person or an agent proposes the action.
Every run leaves an audit trail
What was true, what was predicted, what was proposed, what was chosen, and what happened. Forecasts are graded against what followed, so a model that drifts shows up as misses rather than as silence.
That record is what the next round of improvements is judged against.
The same agent, with and without the platform.
Ergodic evaluated its approach on four public benchmarks, each with its method and caveats published, run with these calls available and without them.
Live incidents repaired by smaller models, 47% to 100%
Root cause named first, 79.4% to 90.6%
Money at the end of a trading month
Cash after 91 days running a simulated company
One result is worth reading twice. Handing the agent a forecast it could consult changed little. Putting the forecast inside the decision changed the outcome by 28%. Access is not use, which is why where these calls sit in your loop matters as much as having them.
Give your agents a model to simulate against.
Bring the agent you are building, or the one already running. We will show you the calls it would make against a model of your operation, and what the benchmarks say that changes.
Talk To An Expert