Plan
The instruction becomes an ordered step list, each one naming the system it will touch and what counts as done. You can read the plan before a single call is made.
CapabilitiesAgents & Automation
Agents that finish the work, inside the boundary.
An agent holds a goal, not a script. It plans the steps, calls your systems, reads what actually came back, and adjusts, running until the work is done or a limit you set stops it. Every call it makes is logged with its arguments and its result.
Agent run · Q3 reconciliation
The run loop
It plans, calls a tool, reads the result, and corrects itself.
A Barie agent is not a flow chart with a language model in one box. It keeps the goal, chooses the next call, and decides the step after that from what actually came back, which is why a run survives a rate limit, an empty result or a schema that changed last quarter.
Call
Each step is a real call: a connector query, a named skill, a browser session, or code the agent writes and runs in a sandbox.
Check
The result is tested against the step’s success condition. An empty page, a 429 or a row count that does not reconcile re-plans the run instead of being passed downstream.
Write & report
The final action lands in the system of record. The run log lists every call, its arguments, its result and which model made each decision.
What an agent can reach for
Six kinds of call, one run.
An agent is only as useful as what it can actually do. These are the call types available inside a single run, mixed freely as the plan requires.
- Connector calls. Read and write across 100,000+ systems (ERP, CRM, warehouse, ticketing) over their own APIs. See Connectors.
- Skills. 270+ named procedures with fixed inputs and a known output shape, so a recurring job runs the same way every time. See Skills.
- Browser sessions. For the systems with no API at all, the agent drives a real browser under your own logins. See Browser Use.
- Code it writes and runs. When no tool fits, the agent writes code and executes it in a sandbox inside your environment, then reads its own output. The same engine works inside your editor and opens the pull request, see Coding Agent.
- Search, internal and live. The same run can query your resolved graph and the open web, and reconcile the two. See Deep Research.
- Sub-agents. A long run can fan independent branches out in parallel and merge the results, rather than walking one step at a time.
Where you stay in control
Autonomy with a hand on the brake.
Every one of these is set per agent, and an agent cannot loosen its own settings mid-run.
- Approval gates. Mark a step, a system or a class of write as requiring a human OK. The run pauses there and waits rather than guessing.
- Dry runs. Execute the plan with every write suppressed, and read the exact changes it would have made before you let it go.
- Step, time and cost ceilings. A run that hits its limit stops and reports instead of looping until someone notices.
- Credentials it cannot widen. Every call executes as the person who started the run, so an agent reaches exactly what that person could already open. No service account with the union of everyone’s access.
- Stop, mid-flight. Halt a running agent from the console. The step in flight completes or reverses, and the log records exactly where it was stopped.
- Triggers you choose. Start a run from chat, on a schedule, or from an event in a source system. See Scheduled Tasks.
When a step fails
The interesting part of automation is the failure path.
Demos show the happy path. What decides whether you can leave a job unattended is what happens on the run where the third step returns nothing.
Transient errors retry
Timeouts, rate limits and locked records retry with backoff. Only a step that keeps failing escalates.
A dead end re-plans
If a step cannot succeed as written, the agent plans a different route to the same goal rather than aborting the run.
Applied writes reverse
Each write step carries its own compensating action, so a run that gives up halfway does not leave your systems half-updated.
Any run replays
Re-run from any step with the original inputs to see what the agent saw. The owner gets the log attached to the failure notice.
Own your intelligence.
Give us a workflow your team repeats every week and watch an agent carry it out end to end.