Agent contracts, execution lifecycle, and bounded autonomy.
Agent Runtime
An agent in AgentWorld is a bounded worker.
It is not a free-form personality. It is an executable role with a contract.
Agent contract
Every agent should define at least these fields:
identity
role
objective
input schema
output schema
memory scope
allowed tools
spend and token budget
escalation conditions
handoff targets
review requirements
The runtime uses this contract before any model call happens.
Example agent definition
The important property is boundedness. A well-formed agent should be easy to reason about from its contract alone.
Runtime lifecycle
1
Receive task
The runtime receives a task envelope with context, policy, and memory scope.
2
Load state
The agent reads only the memory objects and artifacts allowed by policy.
3
Plan next action
The model proposes the next action. The runtime validates that action against tool scope and policy.
4
Execute bounded side effect
A tool call, handoff, or artifact write is executed and recorded.
5
Persist transition
The system stores the new state, emitted events, and operator-visible logs.
6
Exit or continue
The run completes, waits for approval, retries, or hands off to another agent.
Deterministic wrapper, probabilistic core
The model remains probabilistic. The wrapper around it should not.
The runtime should make deterministic decisions about:
what state is visible
which tools are callable
how outputs are validated
which actions need approval
how retries are handled
when the run must stop
This is the main safety property of the system.
Output validation
Every meaningful output should be validated before it becomes business state.
That validation can include schema checks, policy checks, score thresholds, and downstream confirmations. Free-form text is allowed as an artifact. It should not become authoritative state unless it passes a typed gate.
Handoffs
A handoff is not just a message.
A proper handoff includes:
source agent
target agent
task ID
reason for handoff
summarized context
required artifacts
policy context
expected completion condition
This makes agent collaboration explicit and auditable.
Retries and recovery
The runtime should classify failures before retrying.
Transient failures include network errors, model rate limits, and temporary tool outages.
Deterministic failures include invalid input, blocked policy, missing approvals, and schema mismatch.
Risk failures include suspicious destinations, policy drift, and out-of-budget execution.
Only transient failures should retry automatically by default.
Good runtime behavior
A good agent run is short, bounded, and composable. It should emit small state transitions, not giant hidden reasoning chains. The system should prefer many inspectable steps over one opaque jump.
Do not grant an agent a tool because it might be useful. Grant it only because the role cannot complete its task without it.
