Missions & Command Center

Overview

When you submit a complex query — one that requires multiple distinct research steps, analyses, or deliverables — Praxiom automatically detects the complexity and proposes a Mission: a coordinated multi-agent execution plan.

A Mission decomposes your query into a DAG (Directed Acyclic Graph) of subtasks, runs them in parallel where dependencies allow, accumulates context across steps, and synthesises a final summary.

---

How Missions Work

Complexity Classification

Every chat query is classified before execution:

Level	Definition	What Happens
Simple	Single-action query	Executes directly as a single agent
Medium	Multi-step, single agent	Runs in one extended session
Complex	Requires decomposition	Proposes a Mission

The classifier uses a fast Haiku call (~200ms) running in parallel with intent inference, so classification adds no perceived latency.

Mission Proposal

When a complex query is detected, Praxiom emits a mission_proposed SSE event before any agent runs. The chat UI shows a Mission Proposal card with:

• The list of subtasks and their dependencies
• The workflow type for each subtask (synthesis, recommendation, drafting, chat)
• Agent count and maximum parallelism
• Estimated credit cost

You have three choices:

1. Launch — Execute as a multi-agent mission
2. Single agent — Run the query without decomposition
3. Edit — Adjust the proposal before launching

Execution (DAG Layers)

Once launched, subtasks execute in topological layers:

Layer 0:  Subtask A ──────────────────────────┐
          Subtask B ──────────────────────────┤  (parallel)
                                               ↓
Layer 1:  Subtask C (depends on A, B) ─────────┐
                                               ↓
Layer 2:  Subtask D ──────────────────────────┤  (parallel)
          Subtask E ──────────────────────────┘

Subtasks in the same layer execute concurrently via asyncio.gather(). Each subsequent subtask receives an injected context prefix listing artifacts created by prior steps:

[Context from prior steps (2 of 5 completed):
 Step 1 (Market Research) created 4 insights, 2 documents.
 Step 2 (Competitor Analysis) created 3 recommendations.]

<your subtask prompt>

Mission Status

Status	Meaning
pending	Proposed, awaiting launch
running	Subtasks executing
completed	All subtasks finished successfully
partially_failed	Some subtasks failed, others completed
failed	Mission could not complete
cancelled	Cancelled by user

---

Real-Time Controls

While a mission runs, you can intervene on individual agents:

Redirect a Running Agent

Change an agent's directive mid-execution without stopping the mission:

1. Find the agent in the Execution Panel
2. Click Redirect (↩)
3. Enter a new directive
4. The agent's prompt is updated — the new instructions take effect at the next execution checkpoint

Redirect history is preserved per-subtask for auditing.

Stop an Agent

Cancel an individual agent (e.g. if it's no longer needed). Agents that depend on a stopped agent will not execute.

Spawn a New Agent

Add a new agent to an in-progress mission:

1. Click + Add agent to mission in the Execution Panel
2. Set a title and prompt
3. Optionally declare dependencies on existing subtask indices
4. The new agent is queued and executes once its dependencies complete

---

Command Center

Access the Command Center from the sidebar navigation to see a fleet view of all missions across your workspace.

Layout

• Active — Missions in pending or running state with live status indicators
• Recent — Completed, failed, or cancelled missions

Each Mission Card shows:

• Status dot (pulsing for running)
• Title and agent count
• Elapsed or total execution time
• Overall quality score (RQS)
• Artifact summary (insights, documents, recommendations created)

Click any mission card to navigate to the originating conversation.

---

Mission Results

After completion, a Mission Results panel appears in the conversation showing:

• Total execution duration
• Completion summary (synthesis of all subtask outputs)
• Artifact pills (total counts by type)
• Overall quality score and total credits consumed
• Per-agent breakdown with individual quality scores and artifact counts

---

Quality & Reliability

Each agent within a mission goes through the same quality verification as standalone agents:

1. Algorithmic contract — checks minimum artifact counts and citation requirements for the workflow type
2. Independent verification — a separate Haiku call evaluates output quality (0–1 score)
3. Self-healing retry — if quality falls below threshold, the agent retries with a targeted healing prompt (up to 3 attempts)

The mission's overall_rqs is the aggregate quality score across all subtasks.

---

Credits

Credit consumption for missions:

• Each subtask is billed independently using the cost-proportional credit system
• A pre-flight check runs before launch — if estimated mission cost exceeds your credit balance, the mission is blocked with an error message
• The mission record stores overall_credits (sum of all subtask costs)
• A credit floor of 0.25 applies to failed subtasks; cancelled subtasks are charged at 50%

---

Polling & Real-Time Updates

The Command Center polls /api/missions?workspace_id=... every 10 seconds to refresh mission state. Within an active conversation, SSE events provide immediate updates as agents start and complete.

See Streaming SSE for the full list of mission-related events.

What's Next

• Missions API Reference — full endpoint and model documentation
• Streaming SSE — mission SSE event payloads
• Billing & Plans — how mission credits are calculated