AI Agents

You stopped counting how many times you opened the terminal to check if the agent finished. Or how many times you had to re-explain the context because it lost track. Or how many times you waited 20 minutes only to discover it failed at step 3.

This is working with agents today: constant babysitting.

Almirant changes the model. Assign work to an agent and walk away. The agent executes, reports, and you review when it is ready. Multiple agents working in parallel with dependency awareness. The right model for the right task. No checking six terminals.

What is an agent

An agent is an artificial intelligence worker that receives a work item and executes it end to end. It does not suggest -- it executes:

• Reads the project context (stack, repositories, documentation).
• Analyzes the assigned work item (description, acceptance criteria).
• Executes the task autonomously.
• Reports the result with evidence (code, logs, summary).

Available providers

Almirant supports multiple agent providers. Each has its strengths:

Provider	Engine	Ideal for
claude-code	Anthropic Claude	Deep reasoning, complex code, architecture
codex	OpenAI Codex CLI	Autonomous task execution, scripting, automation
zipu	Lightweight agent	Quick tasks, point fixes, simple changes

The system automatically selects the optimal provider based on task type, or you can choose manually.

Runner skills

Agents execute work through specialized skills:

Skill	Description
implement	Implements functional code from the work item description
validate	Validates that code meets acceptance criteria
nightly-fix	Automatically fixes errors detected in nightly builds
document	Generates technical documentation for implemented code
review	Reviews code or pull requests looking for issues
record-video	Records a demo video of the implemented functionality

Jobs: the unit of work

Each time you assign an agent to a work item, a job is created. A job is the execution unit that represents a specific agent task.

Job states

queued --> running --> completed
                  \--> failed
                  \--> cancelled
                  \--> waiting_for_input

State	Meaning
Queued	The job is in queue waiting to be processed
Running	The agent is actively executing the task
Completed	The task completed successfully with a result
Failed	The task failed due to an error (timeout, API error, etc.)
Cancelled	The job was manually cancelled
Waiting for input	The agent needs clarification before continuing

Viewing job status

From the work item detail you can see all associated jobs, including:

• The provider that executed it.
• The current job state.
• The execution duration.
• The result or error produced.
• Real-time logs during execution.

Workflow

The process is straightforward:

1. Select a work item -- Go to the work item you want to execute.
2. Assign an agent -- Choose the appropriate provider and skill.
3. The agent executes -- The job transitions to running state and the agent works autonomously.
4. Review the result -- Once completed, review the output (generated code, created PR, documentation).
5. Accept or iterate -- If the result is satisfactory, merge the PR. If it needs adjustments, launch a new job with additional instructions.

Multiple agents in parallel

You can launch multiple agents simultaneously on different work items. The system manages dependencies and priorities automatically.

Cost tracking

Each agent execution consumes tokens from the AI provider. Almirant records this information:

Metric	Description
Tokens consumed	Total number of tokens (input + output) used
Estimated cost	Approximate cost in USD based on provider pricing
Duration	Total job execution time

Check accumulated costs from the organization settings. See Providers and Quotas for more details.

Prerequisites

To use agents you need:

• A configured AI provider with a valid API key.
• Available quota in the organization.
• A project with at least one work item created.

warning

If your organization's quota is exhausted, jobs will remain in queued state until the quota is renewed or expanded. See Providers and Quotas to manage your quota.

For Developers

MCP Tools

The following tools are available via MCP to interact with agents:

Tool	Description	Main parameters
complete_ai_task	Marks an AI task as completed with result	taskId, result, tokensUsed
record_ai_session	Records an agent work session	projectId, prompt, response, tokensUsed
get_ai_sessions	Gets the AI session history	projectId, limit

Example: Completing an agent task

Tool: complete_ai_task
Parameters:
  taskId: "work-item-uuid"
  result: "Implementation completed. 3 files created: user-service.ts, user-repository.ts, user-types.ts"
  tokensUsed: 4500

Example: Querying project sessions

Tool: get_ai_sessions
Parameters:
  projectId: "project-uuid"
  limit: 10

Returns the last 10 sessions with details on tokens consumed, prompts, and responses.