Find the work
Identify the task and whether an agent can help.
Agent Experience
Agents need clear instructions. People need control and evidence.
Siftable connects the task, relevant context, permissions, execution status, verification, and recovery steps. The agent can see what to do and what it may change. The person can see what happened and why.
A complete delegation · 10 stages
A complete delegation has ten stages.
The work starts with identifying and defining the task. It ends after the result is checked, failures are handled, and useful changes are reviewed for next time.
Gather context
Collect the relevant information and flag gaps or conflicts.
Define the result
State the outcome, constraints, and acceptance criteria.
Make a plan
List the steps, dependencies, and expected changes.
Set permissions
Define permitted actions, resources, limits, and expiration.
Do the work
Perform the approved actions and record what changed.
Track progress
Show blockers, changes, elapsed time, and cost.
Check the result
Compare the result with the acceptance criteria.
Repair
Retry, roll back, revise the plan, or ask for help.
Review and learn
Decide what should be reused or changed next time.
Three linked records
Planning, delegated work, and execution history need separate records.
The records share context and identifiers, but each one answers a different question.
Human task
Defines the goal, owner, priority, dependencies, and decisions that remain with a person.
Agent work item
Defines the scope an agent can claim: instructions, acceptance criteria, allowed actions, write boundaries, and verification.
Execution run
Records one attempt: actor, model and runtime, granted authority, actions, errors, receipts, cost, and timing.
Six required views
Each stage should show the information needed for the next decision.
Siftable derives these views from shared state and presents them when needed. Users do not maintain six separate forms.
Task
What needs to be done, who owns it, and how will success be judged?
Context
Which facts, sources, gaps, and conflicts matter to this work?
Capability
Which actions are available, and what effects or failure modes do they have?
Authority
What may this actor read, change, spend, send, or publish?
Outcome
What changed, which checks passed, and what remains unresolved?
Learning
What should be reused, corrected, or reviewed before next time?
Proportionate control
Match the control to the risk of the action.
Reversible, internal work can move quickly. External, expensive, sensitive, or irreversible actions need stronger limits and approval.
| Class | Action | Default control |
|---|---|---|
| C0 | Inspect | Search, summarize, and calculate. Log access; no approval. |
| C1 | Prepare | Draft, classify, and propose. Preview or sample; easy discard. |
| C2 | Bounded internal change | Explicit scope, lease, verification, and receipt. |
| C3 | Consequential commitment | Human approval at the send, publish, spend, or production boundary. |
| C4 | Exceptional or irreversible | Stronger identity checks, a policy or second-person gate, and a documented recovery plan. |
Outcome assurance
Completion requires evidence.
Check the result at the level the task promised: tests, diffs, receipts, reconciled records, rendered artifacts, source citations, or an explicit human decision.
Failure and recovery
Every failure state should include a valid next action.
Blocked work, stale context, rejected outcomes, partial completion, and expired grants should lead to a clear recovery choice.
Retry safely
Retry the action without duplicating a completed change.
Refresh and replan
Update stale context while preserving the earlier evidence.
Change scope or escalate
Revise the scope explicitly or send the decision to the right person.
Roll back or create follow-up work
Reverse what can be reversed and keep the unresolved work visible.
How Siftable represents the work
Siftable links three records in one shared graph.
Human task
Stores the goal, owner, priority, dependencies, and decisions that remain with a person.
sift work
Stores the claimable scope, allowed actions, acceptance criteria, and verification plan.
Execution run
Stores what happened during one attempt, including actions, errors, receipts, timing, and managed cost.
Because the records share identifiers and context, a reviewer can trace the work from the original goal to the final evidence.
Connect an agent
Set up an agent with shared context and a work queue with clear limits.
Use MCP or the Siftable CLI to give an agent the context and actions needed for a specific task. The work item keeps the result, verification, and source references together.