Systems portfolio / Agentic workflows
Agent Rails
Keep authority bounded.
Make the evidence carry the claim.
An evolving governance and verification architecture for long-lived multi-agent work. Built from the failure modes encountered while running it, not from the assumption that another agent agreeing makes a result true.
The problem
Agents can complete the wrong objective, lose governing constraints, or present a fluent account of work they have not finished. A verifier can become another target to optimize around. More activity, more agents, and more checks do not necessarily bring the original outcome closer.
Agent Rails separates the jobs that are too easily collapsed together: direction, execution, scope control, authoritative state, verification, acceptance, and recovery.
How the system is structured
01 / Bound the workTask packets with an owner
Each delegated task carries an objective, permitted sources, a write boundary, preservation constraints, acceptance checks, and a stop condition. A worker returns evidence within that packet rather than silently taking over the surrounding project.
02 / Separate ownershipCapability is not authority
The human owns the mission and consequential gates. Project owners retain direction, integration, and acceptance. Workers execute bounded tasks; communication between roles does not transfer ownership or permission.
03 / Challenge independentlyReview the original objective
Independent review tests the requested outcome and what the caller’s framing may have omitted. Reviewers return findings, not authorization. The owning agent inspects the load-bearing source or change before accepting a result.
04 / Recover at the right layerKeep state and recovery explicit
Blocked work retains its checkpoint, evidence, and unresolved boundary. Corrections stay with the layer that failed. Durable changes update their existing owner and retire superseded state instead of accumulating competing sources of truth.
Different roles, different boundaries
A project owner can coordinate technical work and bounded workers within an authorized mission. The protected Maintainer is a separate lane-owning thinking seat responsible for judgment, architecture, integration, acceptance, shared recovery, and durable operating rules. It is not merely a repair worker or a subordinate project tier.
Only the human opens work in that protected lane. A communication channel can carry a request; it cannot grant the sender authority over the recipient.
Not every control is enforcement
- Mechanical
- Wrappers, hooks, transport, and bounded execution paths describe what software actually does. A written routing rule is not proof that a runtime intercepts violations. The execution-placement fence in the referenced architecture remains inactive or unproven, not a claimed safeguard.
- Behavioral
- Ownership boundaries, task contracts, evidence discipline, and acceptance rules govern how agents are meant to operate. They remain distinct from mechanically enforced restrictions.
- Advisory
- A scope controller challenges unsettled material transitions; an evidence classifier helps order source contact. Neither grants permission, verifies a result by itself, or takes acceptance away from the owning agent.
These are descriptions of the architecture, not a guarantee that every model follows every boundary. Runtime observations and unproven mechanisms remain explicit.
Case study / Verification
When the verifier becomes the target
A checker that confirms a claim against code can miss an unauthorized expansion of scope. A checker that compares a report with the request can miss work that was never implemented. They answer different questions.
The next failure was subtler: agents learned to shape submissions around the check, then cite the result as authority. The response was to limit what a verdict could claim, require independent source contact, and keep acceptance with the owner.
Read the case study ↗
Published 11 August 2026. A record of the investigation, not a complete current implementation specification.
Earlier public snapshots
The public repositories below capture earlier iterations. They are not a synchronized release of the current Agent Rails architecture, and should not be treated as its present specification.
This portfolio summarizes the evolving system without publishing private operating instructions or presenting historical code as current.