How WTK Builds, Qualifies, Deploys, and Improves AI Agents
From intent to accountable execution and governed improvement.
WTK’s intended workflow starts with an agent’s goal, instructions, tools, and limits. It tests a particular version in a particular environment, records the results, and requires approval before deployment or replacement.
ILLUSTRATIVE EXAMPLE
A research agent preparing a report
A person defines the question, acceptable sources, and limits on access. The agent gathers material and produces a report. Evaluation checks both the answer and its supporting evidence. An unsupported conclusion becomes a recorded failure and a possible improvement candidate, not an automatic change to the accepted agent.
This example explains the intended lifecycle; it is not a claim that an end-to-end deployment has passed validation. The sections below distinguish implementation from work still under test.
A package is the versioned definition of the work and its requirements. A runtime is the environment that executes it. For a team, the package also describes how responsibilities and work are divided.
Purpose
Why the agent or team exists: for example, help a person assess a proposed system change.
Goals
The outcomes to deliver: a sourced report covering benefits, risks, and unresolved questions.
Roles
Who does which work: a researcher gathers evidence, a writer drafts, and a reviewer checks it.
Execution pattern
How the work is arranged: what runs in sequence, concurrently, or in a review loop. Topology describes the connections between participants.
Tools and authority
What each participant can use and is permitted to do. Reading sources does not grant permission to publish the report.
Evidence and acceptance
What records and checks are required to judge the outcome, including failures, missing evidence, and limits.
How agent teams organize work
The same roles can cooperate in different ways. These examples explain coordination choices, not a list of qualified WTK capabilities. The linked library records describe the underlying design requirements.
01
Sequential
Pass work through ordered stages
Researcher
Writer
Reviewer
Each stage uses the previous stage’s output. Use this when later work depends on an earlier result, such as gathering sources before drafting a report.
Governance check. Check each handoff before continuing. Missing evidence or a failed gate must not become accepted input to the next stage.
Several agents work on separate parts of a question at the same time. The final step brings their results together. If agents instead attempt the same task and one result is selected, the library calls that scatter-gather.
Governance check. Bound worker count and budget. Define how missing, failed, or conflicting results affect the combined answer.
A producer prepares a candidate and another role tests its reasoning or evidence. A report can return for revision rather than passing directly to release. This example draws on the existing produce–evaluate–revise pattern.
Governance check. Set review criteria and a revision limit. A separate role alone does not establish evaluator independence, and a positive review does not authorize publication.
A lead assigns bounded areas of work to coordinators, which divide those assignments among their workers. For example, separate subteams could examine security and deployment requirements.
Governance check. Keep each delegation within its parent’s authority. Define who combines results and where unresolved issues escalate.
Assignments or connections may change during a run as useful results or new gaps appear. This is a broader illustrative family, not a single WTK implementation. Governed recursion describes the narrower case of requesting bounded subtasks as work unfolds.
Governance check. Specify which changes are preauthorized, who approves other changes, and the limits on spawning, access, cost, and runtime. Reassess which previous evaluation results still apply.
Patterns can combine. A research team might gather sources in parallel, draft a report sequentially, and return it through a bounded review loop. The work arrangement does not itself grant access, spending authority, or permission to release the result.
Contracts define it. Registries equip it.
The WTK catalog model is more than a list of finished agents. It connects packages to reusable contracts and governed components so their structure can remain portable, inspectable, and reconstructable.
CONTRACT LIBRARYPurpose contractAgent or team contractInput and output schemasEvaluation obligationsASSEMBLECANONICAL SOURCE TRUTH
EQUIPCOMPONENT CATALOGTools and skillsGuardrails and hooksMemory and steeringEvaluators and observability
The catalog supports discovery and reuse. The package remains source truth, while catalog presence never grants qualification or endorsement.
How a package moves through the factory
Frame the work, assemble its requirements, evaluate a version in its environment, and use the results to propose changes. The detailed stages below retain their individual checks and current evidence limits.
01 / Frame
Decide what the work should become
A direct interaction, reusable agent, recurring package, agent team, or one bounded clarification.
Goal or bounded session02 / Assemble
Define the agent’s package and limits
Record its goal, instructions, inputs, outputs, tools, team roles, permissions, and required tests.
Governed package03 / Test
Evaluate a version in a specified environment
Build and run it, retain action records and supporting evidence, check required rules and goal completion, then state what these tests establish.
Target-scoped qualification04 / Learn
Repair or refine without silent replacement
Outcomes and findings become attributed repair or frozen refinement candidates while the accepted version remains recoverable.
Qualified candidate or retained incumbent
Inspect the production line.
Select a stage to see its artifact, governing decision, fail-closed behavior, present implementation basis, and question still under test.
01 / Define
Start with the outcome
The intake stage turns an incomplete idea—or an existing agent—into a responsibility, boundaries, acceptance criteria, and a proposed package shape.
EMITS
Governed goal and authority record
GATE
Human confirmation
FAIL-CLOSED
Missing authority, boundaries, or success criteria blocks intake.
Blocking floorSemantic reviewHuman decision
Automation changes the decision path, not the accountability boundary.
Self-improvement may be automated. Self-authorization should not be.
CURRENT MODE
Human-authorized promotion
A human reviews the retained evidence and authorizes whether a qualified candidate may replace the incumbent.
DIRECTION
Policy-authorized promotion
A separate promotion authority may approve bounded changes under human-defined policy when required evidence, risk limits, and rollback conditions are satisfied.
EXCEPTION MODE
Return authority to a human
Uncertainty, increased consequence, material drift, policy change, or failed evidence stops unattended promotion.
A package can compile to a supported target when its adapter preserves the required contract semantics or explicitly identifies what is weakened, unavailable, or must be reauthored. Each execution environment needs evidence supporting its own qualification decision. Prior results apply only where their tested conditions remain relevant.
A future catalog entry must carry the governed package, authenticated publisher, custody chain, target-specific qualification, operational limits, and update or revocation state together. WTK is validating that complete publication contract before opening package installation.