THE AGENTIC LIFE CYCLE
Experience becomes useful
when learning is accountable.
The research ambition is an agent that develops a recognisable working history while remaining inspectable, correctable and under the authority of its owner.
01 / INSTANTIATEBegin with a defined role.
Choose a task, an initial specification and approved tools. Check the dependencies and permissions before an instance starts. A useful first assignment is narrow enough that a reviewer can recognise success and spot a dangerous mistake.
02 / EXPERIENCECapture evidence, not everything.
A result can produce a task record, a preference or a candidate lesson. Keep provenance and uncertainty with the memory. Separate one person’s private context from shared knowledge and define expiry, correction and deletion.
03 / REFLECTTurn failure into a proposal.
The agent can compare a result with the objective and suggest an altered prompt, a retrieval rule or a more reliable sequence of tools. Reflection is a source of hypotheses. Confident self-assessment does not replace external task evidence.
04 / EVALUATEMake the test independent.
Evaluate the candidate in a sandbox against held-out tasks and known failure cases. Keep a baseline. Check whether improvement on one task harms another, exposes sensitive data or increases resource use beyond the agreed budget.
05 / APPROVEPromote a specific version.
Record the exact changed configuration, the result, its parent and the approving authority. Low-risk adjustments can follow a pre-approved policy; changes to privileges, sensitive access or external commitments need the appropriate owner decision.
06 / OPERATEWatch for drift. Keep a return path.
A model update, changed tool or new task distribution can invalidate old results. Monitor the chosen indicators and stop on defined failures. Roll back the active version while retaining the evidence needed to understand what happened.
Memory is not
model training.
Retrieving an approved memory supplies context at execution time. Altering a workflow changes how tasks are performed. Fine-tuning changes model parameters and introduces a different training, licensing and evaluation problem. These mechanisms should have different version records and approval policies.
A proposed agent can become distinct through scoped memory and selected workflows without training a private foundation model. That distinction keeps the architecture practical and makes each claim easier to test.
THREE ILLUSTRATIVE PATHS
The same seed can
grow in different directions.
The research partner
Learns a source-quality rubric from approved examples. Proposes a better citation workflow. Earns promotion by identifying unsupported claims on held-out documents, with evidence the reviewer can inspect.
The operations assistant
Recognises recurring exceptions and proposes a more precise escalation rule. Tests in a replay environment. It can improve routing without obtaining authority to make payments or change production access.
The learning companion
Adapts explanations to consented preferences and observed exercises. Tests whether a learner can apply the idea, rather than merely click through a lesson. Personal history remains scoped to that learner.
Many agents, distinct responsibilities.
Distributed work could place execution near authorised data, use specialist models for selected tasks and call larger models when necessary. An orchestrator should preserve task ownership, provenance and tool boundaries across the handoffs. More participants also create more points where unreliable output can become someone else’s input.
The recovered Promptus and Ubiquita discussions motivate contributed computation and shared resources. They do not establish a currently deployed Promptinals worker network. The research challenge is to specify trustworthy scheduling, workload verification, data isolation and incentives alongside the agent’s own learning loop.
Discuss a use case ↗