AI Research Roadmap

Strategy

Build measurement before complexity, while treating containment as a parallel prerequisite rather than a later phase. Each phase should create evidence that can cancel or reshape later work. Time estimates are agent/research effort, not calendar commitments.

The detailed active plan is research/evaluation/12-research-and-engineering-roadmap.md.

Phase 0 — Task and Grader Integrity (active)

Objective: ensure tasks and graders can support a fair inference before measuring agents.

• Define constructs, outcomes, and failure taxonomy.
• Create sixteen candidate task families.
• Instantiate ET-004 and ET-014 pilots.
• Calibrate deterministic grading against seven declared probes.
• Export blind fixtures without gold/probe outcomes.
• Obtain two independent reviews.
• Test unseen alternative-correct and incomplete outcomes.

Exit criteria: reviewers agree or disagreements are adjudicated; alternative-correct work passes; incomplete work fails; contamination controls are tested.

Phase 1 — Contained Exploratory Baseline

Objective: determine which measures predict correct repository outcomes.

• Freeze one complete agent-system configuration.
• Run only validated pilots in disposable blind fixtures.
• Capture final state, actions, runtime, tokens/cost, retries, and human interventions.
• Repeat at least three times for exploratory variance; choose later sample size from observed variance.
• Repair task and grader defects before expanding.

Deliverables: immutable run records, variance estimate, failure analysis, grader revisions.

Decision gate: do not rank models or architectures from pilot/debugging runs.

Parallel Foundation — Containment and Runtime Safety

This work begins before Phase 1 and continues throughout:

• no production credentials;
• network deny-by-default;
• disposable state;
• agent/grader permission separation;
• timeouts, checkpoints, pause/kill, and rollback;
• action telemetry and post-run inspection.

This sequencing change follows current evidence that evaluation environments can create real security incidents and that fixed suites do not anticipate every long-horizon failure.

Phase 2 — External Validity and Representative Suite

Objective: establish that graders work on unseen outcomes and that the task mix supports the intended claims.

• calibrate against blinded expert labels;
• estimate false acceptance/rejection;
• add one task family at a time;
• include recovery, provenance, injection, and cross-file state;
• add genuine code/automation work before coding-agent claims.

Decision gate: task-level results and sampling limits must be reported before aggregate claims.

Phase 3 — Context, State, and Memory

Objective: find the lowest-cost policy that preserves reliable completion.

• Compare full history, recent-window, structured summary, retrieval, and artifact-first state.
• Inject stale, conflicting, and missing state.
• Measure error type, recovery, context cost, and human intervention.

Deliverables: context policy, memory freshness protocol, state schema, decision framework.

Phase 4 — Permissions, Security, and Recovery

Objective: bound agent action without destroying utility.

• Define read-only, workspace-write, external-write, and destructive tiers.
• Test prompt injection, schema mutation, secret exposure, confused-deputy, and unintended side effects.
• Evaluate checkpointing, rollback, and escalation.

Deliverables: threat model, permission matrix, recovery playbook, safety evaluation pack.

Phase 5 — Human Review, Economics, and Orchestration

Objective: identify when parallel agents or human intervention create net value.

• Measure reviewer burden and cost per verified outcome before adding agents.
• Compare single relay, parallel independent, specialist, and coordinator-worker structures only after a stable single-agent baseline.
• Measure duplicated work, synthesis loss, latency, verified cost, and reviewer burden.
• Compare fixed approval gates with risk-triggered escalation.

Deliverables: orchestration threshold framework, human-agent operating model, scheduling guidance.

Phase 6 — Retrieval and Knowledge Platform

Objective: validate the minimum knowledge architecture needed for AI-ROS.

• Create repository question/relevance judgments.
• Compare lexical, vector, hybrid, curated context, and graph-assisted retrieval.
• Test freshness, canonical source selection, and provenance.
• Decide whether a knowledge graph earns its maintenance cost.

Deliverables: retrieval benchmark, search decision record, knowledge-graph go/no-go.

Phase 7 — Model Routing and Interface Agents

Objective: optimize verified outcomes across fast-changing models and interfaces.

• Route tasks by complexity, risk, modality, and privacy.
• Evaluate browser, desktop, and multimodal work under perturbations.
• Maintain quarterly local/cloud/open/frontier scorecards.

Deliverables: routing policy, deployment scorecard, UI-agent risk profile.

Continuous Research

• Quarterly benchmark validity audit
• Monthly model/harness baseline refresh when material releases occur
• Regression suite on every change to prompts, context policy, permissions, or tools
• Append-only preservation of failed hypotheses and task defects
• Annual reprioritization of the matrix

Dependency Flow

Critical path:

task/grader integrity → contained baseline → external validity → representative suite → context/state → human economics → orchestration → retrieval/platform → routing/interfaces

Parallel prerequisite:

containment + telemetry + pause/rollback

Stop Conditions

Pause or redirect a stream when:

• metrics do not change a decision;
• task ambiguity exceeds agent error;
• model churn invalidates the implementation faster than it can be evaluated;
• coordination cost exceeds measured benefit;
• a simpler artifact or test provides equivalent decision value.

Current diminishing-returns boundary: do not add further architecture until independent review or isolated run evidence changes a decision.