Make It Ready for Real Life

A workflow is not agent-ready because a model can discuss it. The work needs a defined outcome, representative cases, accessible truth, clear authority, observable side effects, recoverable state, and an owner who can judge failure. Agents amplify the operating environment they enter.

Assess four surfaces: the work surface, where tasks and decisions are modeled; the knowledge surface, where current truth and instructions live; the action surface, where tools and permissions change the world; and the operating surface, where runs are observed, evaluated, recovered, and improved.

Score workflow clarity, data quality, context accessibility, tool reliability, permission design, evaluation strength, side-effect reversibility, operating ownership, and user trust separately. One weak dimension can dominate the system. Excellent prompts cannot rescue inaccessible truth; strong models cannot make an irreversible tool safe.

Readiness is job-specific. A system may be ready to draft an internal brief but not publish it, ready to classify expenses but not approve payment, or ready to recommend a schedule but not move another person’s calendar. State the action class and proof layer when declaring readiness.

Design a pit of success: canonical sources are easier to find than stale copies, typed tools are easier than raw shell access, tests run through standard commands, approvals appear with evidence, and receipts are produced by the normal completion path. Safety that depends on exceptional vigilance will erode under speed and scale.

Templates, types, documentation, architecture, plans, tests, and workflows are leverage points because they shape many future actions. Improve the highest layer that can prevent the failure class. A better workflow contract usually compounds further than a clever prompt patch.

Most organizations cannot redesign every system before using agents. Establish a boundary around the new workflow, adapt legacy inputs into typed records, and emit receipts back into existing systems. Add invariants and observability at the seam.

Each clean run can ratchet the environment forward: improve a schema, replace an ambiguous field, add a test, clarify ownership, or retire a duplicate source. Preserve rollback and compatibility while the new path earns trust. Readiness grows through controlled migration, not a declaration that the old world no longer exists.

Long-running systems need more than conversation memory. Preserve intent—the outcome and constraints that should survive many runs. Preserve task state—the current position, dependencies, owner, and next transition. Preserve world state—external facts with source and freshness. Preserve capability state—available workers, tools, credentials, policies, and versions. Preserve evidence state—the events, approvals, artifacts, evaluations, and receipts that explain what happened.

These state classes change at different speeds and require different write permissions. A tool observation may update runtime truth but should not rewrite durable intent. A worker can propose a memory or policy change but may not have authority to approve it. Separate working context from the full historical record.

A long-horizon system repeatedly turns intent into tasks, tasks into worker runs, runs into observations and artifacts, and evidence into updated state. The next cycle begins from that reconciled state—not from a summary invented at the edge of a full context window.

Composition requires durable identities, causal links, event ordering, compaction, checkpointing, and reconciliation with the external world. A new session should be able to answer what the objective is, what already happened, which evidence is authoritative, what is blocked, and what action is safe next.

Agent systems exist inside organizations with people, incentives, policies, vendors, legacy databases, informal workarounds, changing models, and external dependencies. Optimizing one agent can damage the ecology by moving burden to reviewers, hiding work from an accountable team, or making a critical process dependent on a brittle provider.

Evaluate where responsibility moves. Preserve an exit path from vendors and models. Make knowledge and state portable. Include the people doing and receiving the work in evaluation. A system becomes institutional when it can survive turnover, outages, upgrades, policy changes, and disagreement about what good means.

A software factory produces artifacts repeatedly. An institution also preserves purpose, distributes authority, records decisions, trains new participants, resolves exceptions, and learns from consequences. Long-horizon agents require the second model.

Give every persistent workflow a charter, owner, service expectation, authority policy, data boundary, evaluation set, incident path, knowledge maintenance rule, and retirement condition. The system should know not only how to continue but when it should stop being trusted.