reindeer@docs:~$ cat for-agents.md▌
Reindeer for Agents and Builders
Reindeer gives you a place to run your agent, evaluate its behavior, and control how it changes in production.
You bring the agent logic. Reindeer provides the runtime, security controls, integrations, and tools to test and approve changes. Agent managers and builders can evaluate a specific agent revision against real cases before making it active, while operators can provide feedback without directly changing live behavior.
[01]Why builders pick Reindeer
[01.1]Security you don't have to build yourself
- ▸Enterprise private cloud option. Deploy in a dedicated environment under your data governance and access controls.
- ▸Security and compliance. SOC 2 Type 2 and ISO 27001, with controls that support GDPR requirements. Your application's compliance also depends on how you build and operate it.
- ▸Bring your own IdP and SSO. Use your existing identity provider rather than maintaining a parallel user directory.
- ▸Scoped access and managed credentials. Keep credentials in managed secrets and connection profiles, not in prompts or reference documents. Give each agent access only to the systems it needs.
- ▸Auditability. Inspect agent actions, run history, and human interactions when investigating an outcome.
[01.2]Skip the boilerplate
- ▸Use pre-built skills for OCR, document parsing, extraction, classification, and routing, with integrations for systems such as SAP, Oracle, NetSuite, Workday, and ServiceNow.
- ▸Manage reference documents and policies through versioned contexts.
- ▸Inspect runs and cases, collect feedback, and handle human-in-the-loop requests through the platform.
[02]Evaluation and change management
An agent needs to keep working when policies, reference material, or its configuration change. Reindeer treats those changes as agent revisions that can be evaluated and approved before becoming active in production.
Operators can provide feedback in plain English. Changes informed by patterns in that feedback over time are captured in a proposed revision; feedback does not silently rewrite the active agent. Approval applies to the revision, so the agent manager can review the proposed behavior and its evaluation results together.
Use this process when changing an agent:
- 01>Capture the change in a revision. An agent revision is an immutable version of its configuration. Evaluate the exact revision you intend to release. The live configuration only changes when you approve revisions.
- 02>Evaluate against representative cases. Run the revision against a golden dataset, a reusable set of cases for evaluation, or select specific cases. Evaluation criteria define what the results are judged against.
- 03>Review improvements and regressions. Inspect per-case scores, reasoning, errors, and the criteria rollup. Compare candidate and baseline revisions using the same cases and criteria; an aggregate score alone can hide a regression on an important case.
- 04>Approve the revision before activation. Evaluation provides evidence for the release decision. It does not replace approval of the revision that will become active in production.
- 05>Use production feedback to improve the next change. Review exceptions and corrections, add representative failures to future evaluations, and repeat the process for the next proposed revision.
Human-in-the-loop decisions serve a different purpose: an operator can answer a question or approve an action within a particular run. That decision is not approval of a new agent revision. Rather, agent managers configure escalation rules and confidence thresholds for the workflow, and evaluate whether they produce the behavior you need.
[03]Get started
[03.1]Install the CLI
curl -fsSL https://headless.reindeerlabs.ai/install.sh | bash reindeer --version
irm https://headless.reindeerlabs.ai/install.ps1 | iex reindeer --version
[03.2]Log in
reindeer login
For an agent or terminal without an interactive browser, run:
reindeer login --headless --no-prompt
Open the printed URL to complete human authentication. For unattended automation, use the service-account authentication flow in the CLI reference; headless login is not a substitute for service-account credentials.
[03.3]Run and inspect an agent
Use the CLI reference to create a workspace and configure an agent’s contexts, access, and connector subscriptions. Resource schemas evolve, so check reindeer <resource> <verb> --help before constructing a command.
For an agent already configured in your workspace, replace the placeholders below:
reindeer runs create -w <workspace> --agent <agent> --wait \
--input '{"type":"new_run","user_input":"<instruction>","artifacts":[]}'
reindeer runs list -w <workspace> --agent <agent>
reindeer cases list -w <workspace> --agent <agent>Run creation returns an operation. Its completion does not mean the agent has finished; inspect the returned run with reindeer runs get to follow execution.
[03.4]Evaluate a revision
With a golden dataset already configured for the agent:
reindeer agent-revisions list -w <workspace> --agent <agent> reindeer evaluations create -w <workspace> --agent <agent> \ --revision <revision> --wait reindeer evaluations list -w <workspace> --agent <agent> \ --revision <revision> reindeer evaluations get <evaluation-id> -w <workspace> --agent <agent> \ --revision <revision> -o json
Replace <evaluation-id> with the ID of the evaluation you want to inspect. create uses the latest golden dataset by default; pass --cases with full case resource names to evaluate a selected set instead. get returns per-case results and the criteria rollup, while list returns summaries. Running an evaluation does not approve the revision for production.
[04]Core concepts
| concept | what it is |
|---|---|
| Workspace | The scope in which your agents, data, and runs live. |
| Agent | The configured automation, including its context, access, and connector subscriptions. |
| Agent Revision | An immutable version of an agent's configuration, evaluated and approved as a specific change. |
| Artifact | A document or other piece of data the agent can use or produce. |
| Context | A versioned collection of artifact references that supplies policy and reference material to an agent. |
| Connector | An integration that brings external inputs and events into the platform. |
| Case | A persistent record of a business item, including its state and history across runs. |
| Run | One execution of an agent; a case can be handled across multiple runs. |
| Golden Dataset | A reusable set of cases selected for evaluating agent behavior. |
| Evaluation Criteria | The requirements used to judge evaluation results. |
| Evaluation | A test of a specific agent revision against cases, with per-case results and scores against criteria. |
[05]CLI reference
The Reindeer CLI reference covers installation, authentication, resources, execution, evaluations, and profiles. Use it alongside command help and your tenant's OpenAPI schema as the source of truth for programmatic details. This document is the overview.
[06]Architecture, briefly
Reindeer sits on top of your existing stack. It integrates with your enterprise systems and works within your data governance and access controls. Context versions and agent revisions make changes identifiable, evaluations provide evidence about behavior, and revision approval controls what becomes active.
[07]Tips for building well
- ▸Use -w (workspace), -o (output), and --profile consistently across environments.
- ▸Keep routine cases, edge cases, and previous failures in your evaluation coverage.
- ▸Compare revisions on the same cases and criteria, and investigate regressions before approval.
- ▸Test changes in a development workspace before approving them for production.
- ▸Watch runs, cases, and pending interactions after release; turn useful feedback into proposed changes and regression tests.
[08]Who's behind Reindeer
- ▸Built by seasoned executives from Google, Microsoft, and Oracle.
- ▸Backed by Lightspeed Venture Partners.
- ▸Trusted by Fortune 100 companies.
[09]Talk to us
- ▸
- ▸reindeer.ai to reach the team.
// note
This document is written for humans and agents. Use the overview to understand the platform, and follow the CLI reference for programmatic details.