Mission-Oriented Process Model¶

Transforms agents from task-runners into mission-driven processes with LLM-as-judge evaluation, confidence thresholds, escalation, and automatic completion.

When you need it¶

Standard agents run on a trigger, do something, stop. A mission-oriented agent runs until a goal is achieved — checking database migrations, monitoring a deployment, researching a topic until a quality bar is met. The agent accumulates context across invocations and the runtime evaluates whether the mission is complete.

Configuration¶

from promptise.runtime import ProcessConfig, MissionConfig, EscalationTarget

config = ProcessConfig(
    model="openai:gpt-5-mini",
    instructions="Migrate all tables to the new schema.",
    mission=MissionConfig(
        enabled=True,
        objective="Migrate all database tables to v2 schema",
        success_criteria="All tables pass v2 schema validation with zero errors",
        eval_every=3,              # Evaluate every 3 invocations
        confidence_threshold=0.7,  # Escalate below this
        timeout_hours=24,          # Fail after 24h
        max_invocations=50,        # Fail after 50 invocations
        auto_complete=True,        # Stop on success
        eval_model="openai:gpt-5-mini",  # Separate model for eval
        escalation=EscalationTarget(
            webhook_url="https://hooks.slack.com/...",
        ),
    ),
)

Or in a .agent manifest:

name: schema-migrator
model: openai:gpt-5-mini
instructions: Migrate all tables to the new schema.
mission:
  enabled: true
  objective: "Migrate all database tables to v2 schema"
  success_criteria: "All tables pass v2 schema validation"
  eval_every: 3
  confidence_threshold: 0.7
  timeout_hours: 24
  auto_complete: true
  escalation:
    webhook_url: "https://hooks.slack.com/..."

How it works¶

State machine¶

Each mission moves through states:

State	Meaning
`pending`	Mission defined but not yet started
`in_progress`	Agent is actively working
`completed`	Success criteria met
`failed`	Timeout, max invocations, or manual failure

Evaluation cycle¶

Every eval_every invocations, the runtime:

Takes the recent conversation history
Calls a separate LLM (the eval_model) with the objective, success criteria, and conversation
Receives a structured MissionEvaluation: achieved (bool), confidence (0.0–1.0), reasoning, progress summary
If achieved → mission completes, process stops (if auto_complete=True)
If confidence < confidence_threshold → fires escalation, the team can intervene

Context injection¶

Before every invocation the agent sees its mission context in the system prompt:

[Mission] State: in_progress | Objective: Migrate all database tables to v2 schema
Evaluations: 2 | Last confidence: 0.82 | Last summary: 6 of 8 tables migrated

The agent can also call check_mission in open mode to inspect full evaluation history.

API reference¶

`MissionConfig`¶

Field	Type	Default	Description
`enabled`	`bool`	`False`	Enable mission tracking
`objective`	`str`	`""`	What the agent is trying to achieve
`success_criteria`	`str`	`""`	How to judge completion
`eval_model`	`str \\| None`	`None`	LLM for evaluation (defaults to process model)
`eval_every`	`int`	`5`	Evaluate every N invocations
`confidence_threshold`	`float`	`0.7`	Below this → escalate
`timeout_hours`	`float`	`0`	Max hours (0 = no timeout)
`max_invocations`	`int`	`0`	Max invocations (0 = unlimited)
`auto_complete`	`bool`	`True`	Stop process when achieved
`escalation`	`EscalationTarget \\| None`	`None`	Where to notify on low confidence

`MissionTracker`¶

Created automatically when mission.enabled=True.

Method	Description
`.state`	Current `MissionState`
`.should_evaluate()`	Whether eval is due this invocation
`.evaluate(conversation, model)`	Run LLM-as-judge evaluation
`.context_summary()`	One-line string for agent context injection
`.is_timed_out()`	Whether the mission has exceeded `timeout_hours`
`.fail(reason)`	Manually fail the mission

`MissionEvaluation`¶

Returned by .evaluate():

Field	Type	Description
`achieved`	`bool`	Whether the mission is complete
`confidence`	`float`	0.0–1.0 confidence in the assessment
`reasoning`	`str`	Why the evaluator made this judgment
`progress_summary`	`str`	Current progress description
`invocation_number`	`int`	Which invocation this evaluation covers