01
Observe
Record the approved market or system context that the agent received.
Integration overview
Add evidence around your agent, not inside its decision logic.
SOLvival wraps an existing agent workflow with a local, integrity-checked record. The pilot starts with one approved flow, so your team can validate the evidence before expanding anything.
The shape of the work
02
Decide and propose
Record the decision separately from a suggested or attempted action.
03
Finish and verify
Record the outcome and a clear stop reason, then validate the campaign bundle.
A minimal safe example
from services.observability import AgentRun, EventType
run = AgentRun("agent-run", budget_units=50,
max_seconds=30)
observation = {"source": "demo_feed", "head": 1001}
run.emit(EventType.OBSERVATION, observation,
cost_units=1, context=observation)
run.emit(EventType.DECISION,
{"action": "paper_hold"})
run.emit(EventType.ACTION_PROPOSED,
{"action": "paper_buy"})
run.emit(EventType.RESULT,
{"status": "simulated"})
run.emit(EventType.STOP,
{"reason": "complete"})The record stays separate from execution.
This example captures what an agent observed and intended. It contains no wallet handling, transaction signer or order-sending function.
- Events receive sequence, timestamp and integrity-chain data.
- Budget and deadline controls are visible in the resulting campaign.
- The complete bundle contains the record, a report, a manifest and a local dashboard.
MCP workflow evidence
Record a tool-call lifecycle without taking control of the tool.
The included MCP reference pattern captures the available tool, a host-approved decision, a proposed call and the outcome reported by the host. It preserves only a cryptographic digest of the approved arguments, so raw prompts, tokens and other sensitive inputs stay in the host’s own systems.
It is a local reference pattern, not an MCP client: it does not connect to servers, retain credentials or execute tools. A pilot maps this pattern to one approved workflow before any expansion.
What a pilot needs from your team
One owner, one workflow and one approved data source.
We agree the evidence fields, limits and expected stop conditions first. The initial work is local-first and execution-disabled unless the team later approves a different, explicitly scoped environment.