Harnesses make agents operational.
They supply the runtime capabilities through which agents reason, coordinate, use tools, and create outcomes.
WTK / GOVERNED AGENT ENGINEERING
Build governed agents and teams for the harness you choose.
WTK turns explicit goals into canonical governed packages, compiles those packages into supported execution environments, and qualifies what each execution form actually demonstrated.
Goals define the work. Evidence earns the trust.HOW WTK FITS
Harnesses execute. WTK preserves continuity.
Harnesses provide models, tools, memory, sessions, channels, coordination, and persistent runtimes. WTK keeps the agent or team tied to explicit purpose, authority, identity, evaluation, and evidence as it moves between supported execution forms.
WTK makes the work portable and inspectable.
It preserves the goal, delegated authority, contracts, evaluation obligations, evidence requirements, and package identity outside any one runtime.
- 01Explicit goal→
- 02Governed package→
- 03Target projection→
- 04Chosen harness→
- 05Execution evidence→
- 06Scoped qualification→
- 07Controlled improvement
The rigor is itself a hypothesis. WTK must demonstrate that this structure raises a measurable floor rather than merely adding process and cost.
EVIDENCE POSTURE
What current work supports.
A compact public summary of mechanisms and bounded observations. The complete ledger preserves each claim’s evidence basis, limitation, and next falsification test.
5Active research questions
1Additional question in preparation
7 of 8Tracked claims have bounded implementation evidence
0General autonomy guarantees
MechanismImplementedEvidenceBounded observation
Package identity survives current projections.
One canonical package record remains identifiable through the currently observed target projections.
LIMITATIONEquivalent behavior and enforcement across different harnesses remain unproven.MechanismImplementedEvidenceBounded observation
Deterministic failures remain disqualifying.
A favorable semantic review cannot silently upgrade a subject that failed an objective requirement.
LIMITATIONReviewer independence and long-run reliability still require controlled trials.MechanismImplementedEvidenceBounded observation
Improvement cannot silently promote itself.
Candidate comparison, qualification, rollback, and promotion authority remain separate from the improving system.
LIMITATIONSafe long-duration improvement and evaluator drift have not been established.LATEST RESEARCH
Research that challenges the design.
External studies inform WTK experiments and boundaries. They do not become WTK results merely because they support an architectural concern.