Agents

An agent isn’t a prompt. It’s a role on your team.

Name it, choose the model, write the mandate, scope the tools. The same definition runs the same way on every task, in every project — no matter who kicks it off. Define a roster once; let it work in parallel.

Orion
Architect
Claude Sonnet 4.6 · Anthropic
Org
Skills
planningarchitecture
Recent activity
read requirements.md
analyze dependencies
draft architecture.md
map component graph
12 tasks$410
Atlas
Implementer
Claude Sonnet 4.6 · Anthropic
Space
Skills
codingtesting
Recent activity
read handlers.ts
write auth.ts (+142)
run npm test
commit 3 files
28 tasks$1,120
Echo
Reviewer
Claude Sonnet 4.6 · Anthropic
Org
Skills
reviewsecurity
Recent activity
read PR #847 diff
check style rules
flag complexity
approve changes
34 tasks$350
Sentinel
Security Auditor
Claude Sonnet 4.6 · Anthropic
Org
Skills
securitycompliance
Recent activity
scan 847 packages
check OWASP Top 10
audit auth flow
report 0 critical
15 tasks$810
Nova
UI Reviewer
Gemini 2.5 Pro · Google
Space
Skills
designaccessibility
Recent activity
screenshot pages
compare design tokens
check a11y rules
flag regressions
22 tasks$862
Define
your own

An agent is a reusable role — configured once, governed centrally, and run the same way by everyone on your team.

The prompt trap

A clever prompt lives on one person’s laptop. Nobody else can reuse it, nobody can audit it, and the next run drifts the moment the model changes. Good results stay anecdotal — they never become infrastructure.

The agent role

An agent is a named role: a model, a system prompt, a credential, a skill set, and a tool allow-list — versioned in one definition. Assign it to a workflow step and it executes identically every time, with a traceable record on every run.

One definition, everything in it

Model, mandate, credentials, skills, tools — all in one place.

Click an agent to see how each role is configured. Definitions are reusable across spaces and workflows, so a configuration that works gets shared — not rebuilt from a blank page.

Agents5
Atlas
Implementer · Anthropic
Model
Claude Sonnet 4.6
Provider
Anthropic
Credential
anthropic-prod
System Prompt
You are an expert software engineer. Write clean, well-tested production code following the project's conventions. Always include error handling and edge cases...
Skills
TypeScript Standards
Testing Patterns
Tool Allow List
read_filewrite_filerun_terminalsearch_codebase
Compounding capability

AI that compounds across your organization.

When a developer finds the right model, the right prompt, and the right tool configuration for code review — that knowledge shouldn’t live on their laptop. In Spaces it becomes a shared agent definition: available to every team, assignable to any workflow, with a traceable record on every run.

And it gets better over time. Tune the prompt against real execution data. Swap the model when something faster ships. Tighten tool access as you learn what works. Every team using that agent gets the improvement automatically.

Assign roles to steps

The right agent picks up the right step.

Each workflow step names the agent that handles it. When a task reaches implementation, the Implementer claims it. When it reaches review, the Code Reviewer takes over. When it reaches deploy, a human signs off. You decide how much to automate.

Implement
Atlas
Code Review
Echo
UI Review
Nova
MA
Deploy
Marcus Alvarez

Assign agents to steps. Tasks route automatically.

Every run, attributable

See every run in the thread.

When an agent executes a workflow step, Spaces records a structured sequence of events: tool calls, file changes, cost milestones, and completion. Not a casual chat log — an attributable trail tied to the agent definition and the task.

Execution threadcomplete · 0.00s
Atlas
task SPACES-247 · step Implement
14:32:01.042
Agent claimed taskAtlas · SPACES-247
14:32:01.198
read_filesrc/api/handlers.ts
14:32:02.011
search_codebasequery: "auth middleware"
14:32:04.887
write_filesrc/api/auth.ts (+142 lines)
14:32:08.120
write_filesrc/api/__tests__/auth.test.ts (+89 lines)
14:32:11.540
run_terminalnpm test -- auth.test.ts → 7 passed
14:32:15.318
Cost milestone2,847 tokens · $0.024 (cumulative)
14:32:15.401
Step completeImplement · handoff → Code Review
What you configure

Full control over every agent.

Everything you set lives in one definition — visible, auditable, and reusable across your organization. Six controls turn a prompt into governed infrastructure.

system_prompt

System prompts

Define what the agent knows, how it behaves, and the standards it follows — consistent behavior across every task.

model

Model selection

Pick the model and provider per agent — a fast model for routine work, a frontier model for review and security. Whatever fits the role.

skills

Skill tags

Tag agents with skills — coding, review, security, testing. Workflow steps match against skills and route work automatically.

tools

Tool allow-lists

Control which tools each agent can touch. Restrict file access, limit API calls, enforce sandbox boundaries.

scope

Scope control

Org-wide agents available everywhere; Space-scoped agents restricted to specific projects. You set the blast radius.

metrics

Performance tracking

Tasks completed, cost per task, error rate, cycle time. See which agents deliver and which need tuning.

Any model, any provider

Route each role to the model that fits it.

Code review on Claude Sonnet 4.6, visual QA on Gemini 2.5 Pro, a self-hosted endpoint for sensitive work — pick the model per role. Credentials are scoped per agent, so you swap models without rewriting workflows.

Spaces

Managed execution with built-in cost tracking and credentials

Anthropic

Claude Sonnet 4.6 and the full Claude lineup

OpenAI

GPT-4o and newer reasoning models

Google

Gemini 2.5 Pro and the Gemini family

Self-hosted

Llama, Mistral, or any OpenAI-compatible endpoint

Start with one agent. Scale to a team.

Pick a model, write the mandate, scope the tools, assign it to a workflow step. Spaces handles the rest — consistent execution, a full audit trail, and cost tracking per agent and task.