Assist
Draft, retrieve, summarize, classify, or compare. Humans remain fully in the action path.
The Kautilyan Harness
Skills do the work. The Harness makes it governable.
The Harness supplies the right context, invokes the right specialist skills, records what happened, and enforces the review and approval gates appropriate to the risk. It is intentionally thin: expertise belongs in reusable skills, not one giant agent.
Agent proposes. Human approves. Organisation learns.
One controlled run
From evidence to a reviewable business artifact.
The same operating loop can support requirements documents, solution designs, SOPs, test packs, reports, or recurring operational decisions.
01 ScopeDefine the outcome
What must be produced, for whom, and how success will be judged.
02 ContextRetrieve evidence
Use only the permitted sources, examples, policies, and project facts.
03 SkillsApply expertise
Authoring, domain, and analysis skills work within explicit contracts.
04 EvalsReview independently
Separate reviewers test whether the output achieved its purpose.
05 GatesEnforce controls
Deterministic checks and risk-based approvals decide what can move.
06 MemoryRecord learning
Decisions, corrections, and outcomes improve the next run.
The architecture choice
Thin harness. Rich skills. Strong evidence.
Orchestration should route work, assemble permitted context, record state, and enforce gates. It should not contain every domain instruction or pretend to be the source of truth.
Why this matters
- A skill can be improved without rebuilding the entire workflow
- A reviewer can evolve independently from the writer
- Known defects can become deterministic checks
- Context and permissions remain visible and auditable
- The same skill can be reused across approved projects
Governance by consequence
Human attention belongs at decisions, not orchestration chores.
Recommend
Prepare evidence and options. An accountable person approves the consequential choice.
Execute within bounds
Only when authority, rollback, audit trail, and exception handling are explicit.
What compounds
Every failure should improve the system, not create another workaround.
Evidence improves
Missing facts become better data requests, source mappings, and context contracts.
Skills improve
Expert corrections become clearer instructions, examples, rules, and terminology.
Checks improve
Repeatable defects become tests that prevent the same issue from recurring.
Start with one recurring artifact or workflow.
We will map the work, define the acceptance contract, and show where skills, reviewers, deterministic gates, and human approvals belong.