Skip to main content

Agent Workflows

Agent Workflows are BotDojo’s long-running coordination layer for AI and human collaborators—think of them as Jira or Linear boards built for agents. In the codebase they correspond to WorkQueue records, but the feature adds opinionated lifecycle management, assignment controls, and model context sharing so repeated work stays predictable.

Core building blocks

Workflow definitions

  • Capture the data schema for items, including any custom fields that appear in intake forms, list views, or automations.
  • Describe every status in the lifecycle with human- and agent-facing instructions, completion flags, and positioning for the visual designer.
  • List the agents that can be invoked, along with pointers to their backing flows and required model-context resources.

Workflow items

  • Represent individual units of work with their own payload, owner, and progress flags.
  • Track state such as whether an item is ready, running, snoozed, or needs attention so the system can schedule it safely.
  • Preserve version history so errors, retries, and hand-offs can be audited.

Templates and management tooling

  • Templates let you stamp out new workflows that inherit statuses, fields, and agent assignments from a proven definition.
  • The workflow manager surfaces queue-level controls to duplicate, pause, archive, or reactivate work without touching individual items.

Lifecycle and transitions

Statuses outline the path an item follows from creation to completion. Each status can:

  • Belong to a human or AI category, which drives default instructions and UI hints.
  • Restrict which fields are editable, ensuring sensitive data remains read-only when it should.
  • Offer default owners so a transition automatically assigns the next participant.

Transitions declare the allowed moves between statuses. They can require certain fields to be filled, inject validation prompts, or hand ownership to a different agent or person. Because terminal states are flagged explicitly, teams can model multiple “done” outcomes—such as completed, cancelled, or needs review—without resorting to hacks.

Assignments, permissions, and guardrails

An item can be owned by an AI agent, an individual user, or a security group. Permission checks ensure that only the current assignee (or an unassigned item) can update the payload, preventing silent overwrite conflicts. Additional guardrails—such as locking an item while an automation is active or parking it in a snoozed state—stop duplicate work and keep human reviewers in sync with autonomous agents.

Collaboration history and oversight

Every noteworthy event is captured:

  • Status changes log who initiated the move, what changed, and any associated agent session.
  • Notes provide threaded discussion with attachments and preserve the assignee at the moment the comment was added.
  • Combined history feeds let product surfaces reconstruct a timeline grouped by status, making audits and retrospectives straightforward.

Model context resources on every step

Agent Workflows couple ownership with the context required to act. When a status hands control to an agent, the workflow can automatically attach documents, tools, previous outputs, or other model-context resources. Humans who pick up the task see the same curated bundle, so everyone operates with a shared source of truth rather than rebuilding context from scratch.

Resource Manager and cost control

Workflows can be bound to a resource group managed by the Resource Manager. This integration gives you:

  • Budget windows that cap spend over rolling durations (minutes, days, months) and expose current versus maximum cost for dashboards.
  • Circuit breakers that halt expensive inference calls before they spiral.
  • Concurrency limits that throttle how many agent sessions can run in parallel when hardware or budgets are tight.

Because every inference, follow-up flow, and resource lookup inherits the workflow’s resource group, the platform can pause, snooze, or resume work when thresholds are hit—keeping agents productive without surprise bills.

Operating an Agent Workflow

Workflow definitions carry their own lifecycle. They can be marked active, paused, archived, or set to needs attention so operators know whether new items should be created. Bulk actions exist for creating items, reassigning owners, or shifting entire cohorts to a new status, and APIs are available for uploading work, synchronising notes, or exporting history for analytics.

Putting it to work

  1. Model the lifecycle: document the statuses, transitions, and field permissions that mirror your team’s process.
  2. Register the automations: list the AI agents and human groups that own each stage, along with the resources they need.
  3. Configure guardrails: tie the workflow to a resource group with sensible budgets, circuit breakers, and concurrency caps.
  4. Launch and observe: let items flow between humans and agents, monitor the audit trail, and iterate on the template as your process evolves.

The result is a repeatable, auditable orchestration layer where AI assistants and human experts share structured workflows while staying firmly within governance and cost controls.