Define an agent once, with its instructions, skills, and tools. Start it from Slack, a pull request, an alert, or a schedule. Agentrun runs it on Claude Code, Codex, or OpenCode in its own environment, on the model you choose, and keeps a record of what it did and what it cost.
Every run keeps its full trace. Open one to read each step, tool call, and retry.
Anything you can describe in an agent definition can run here. These are the kinds of jobs teams hand over first.
Reads the alert, checks Sentry and recent deploys, names the affected service and a likely cause.
Reviews the diff against your team's conventions, runs the tests, flags what a reviewer should look at first.
Finds the failing test, works out whether the test or the code is wrong, and fixes the right one.
Pulls new product rows, fills in categories and attributes from the text, and flags the ones it is unsure about.
Reads the last 24 hours of alerts, incidents, and open threads, and writes the handoff note the next person needs.
Anything with instructions, a few skills, and the tools it needs. Point it at a repo, a database, or a queue.
The same six steps for every agent, whatever starts it. Here, one incident-triage run.
@triage, checkout p99 latency is over 2s since the 3:50 deploy. Can you look?Example result, posted back to the thread. Every real run keeps its full trace.
Mention the agent in Slack, use a slash command on a pull request, assign it a Linear issue, let a PagerDuty alert start it, put it on a schedule, or call the API. Each run gets its own isolated environment, credentials scoped to that run, and the harness and model you chose. You can talk to a running agent from Slack or the web.
Keep instructions, skills, MCP tools, permissions, and credential references in one versioned definition. Agentrun adapts it for each harness, the software that actually runs your agent, such as Claude Code, Codex, or OpenCode, and tells you what carries over and what needs attention before you switch.
Connect Slack, GitHub, Sourcegraph, Sentry, Datadog, Linear, Zapier, Postgres, or anything else with an MCP server. The connections live in the agent's definition, and every run gets its own short-lived credentials, so a token never outlives the job.
Open any run to read its steps, tool calls, retries, result, and cost. See every agent's runs in one place, and group failures across agents and models so the recurring ones stand out.
Replay tasks from an agent's history across the setups you want to compare. You get pass rates, critical failures, runtime, and cost per successful run, so the decision to move a workflow is a table, not a hunch. Example below, illustrative data.
| Harness | Model | Hosting | Pass rate (95% CI) | Critical failures | p95 time | Cost per successful run |
|---|---|---|---|---|---|---|
| Claude Code | Claude Opus | Model provider | 91% (88 to 94) | 0 | 6m 40s | $2.84 |
| OpenCode | Qwen3-Coder | Self-hosted | 89% (86 to 92) | 0 | 7m 05s | $0.41 |
| Codex | GPT | Model provider | 90% (87 to 93) | 0 | 6m 58s | $2.31 |
| OpenCode | GLM | Self-hosted | Not run | Not run | Not run | Not runGLM did not pass the tool-call compatibility check for this agent. |
86% cheaper per successful run, pass rate two points lower, no critical failures seen. Decide whether a drop of up to five points is acceptable before moving all of it. Other agents get their own comparison.
Illustrative figures. A pass here means the agent named the affected service and produced a diagnosis the reviewer accepted. Cost per successful run includes failed attempts and retries. Self-hosted inference is priced as the GPU time allocated to the run, idle capacity included.
Each plan includes monthly credits for active agent time. One credit is one dollar. Idle environments cost nothing.
Import its existing files, connect its tools, pick where it should start from, and run it. Pick a job you know well enough to judge the result.