Four phases. Pillar maturity over time.
Four phases for the agent runtime infrastructure for regulated production environments. Items move between tiers as the product ships — not as they are planned. The sequence is driven by what regulated enterprise buyers actually need next, not by what makes the feature list look complete.
Estate Foundation
Pillars 1–3 (Runtime, Authority and Accountability, and Lifecycle Control) shipped full. Pillar 4 (Audit-Grade Evidence) shipped foundational. Pillar 5 (Estate Intelligence) in development. Two framework adapters. Tier A deployment.
- Agent registration and manifest — canonical inventory across every framework and team
- Machine identity per agent — never user-borrowed credentials; every action attributable to a specific non-human principal
- Framework adapters (two frameworks) — tool call interception and context propagation; narrow by design
- Durable execution engine — event-sourced, replayable, self-hosted; no vendor control plane dependency
- Lifecycle authority — suspend / cancel / kill / drain / rollback at run, agent, version, or tenant scope
- MCP gateway enforcement — single authorisation point for every tool call; paired with egress control
- Sandboxed execution — process and network isolation per run; graduated profiles
- Token and resource limits — per-agent, per-tenant, per-run; enforced before consumption with circuit breakers
- Cost attribution — per-run cost from registration metadata; no manual tagging
- Tamper-evident audit log — hash-chained, privacy-preserving, compliance team query interface
- Events journal — real-time operational stream with OTel export
- Control Center — Platform Engineer and Security/Compliance views; full API surface
- Tier A — VPC / cloud-isolated — Helm chart deploy; EKS / AKS / GKE; no cogward control plane after deployment
- Tool and datasource governance — the platform governs access; it does not build or maintain connectors
Phase 1 success condition: One regulated enterprise — financial services — running production agents on cogward. Pen test survived. Compliance audit referenced. That reference is the asset everything else is built on.
Production Governance
Pillar 4 (Audit-Grade Evidence) reaches full — DORA / SOC 2 certification, GRC integrations. Pillar 5 (Estate Intelligence) foundation — task status, cost & latency baselines, goal achievement tracking. Evaluation gate, context governance, skills registry. Pillar 1 extended with Tier B deployment.
- Policy hierarchy and pre-execution enforcement — permit, block, escalate before every significant action; never advisory
- Policy versioning and approval workflow — full audit trail on policy changes; change approval before production promotion
- HITL approval queue — ServiceNow, Jira, PagerDuty integrations
- Production readiness checks — pre-promotion gate against policy and resource baselines
- Compliance reporting templates — SOC 2, DORA, and NIST AI RMF as the first pack, driven by first customer requirements
- GRC/SIEM integration depth — structured audit export to ServiceNow, Archer, Vanta, Splunk, and Sentinel
- Minimal production evaluation gate — behaviour validation, policy simulation against test traces, evidence-backed promotion decision
- Forensic trace logging — capture of execution event loops, retrieval contexts, and local memory states during tool calls
- GRC integrations — structured audit log exports to ServiceNow, Archer, and Vanta for real-time risk verification
- Governed memory access layer — secure filtering abstraction over existing enterprise memory infrastructure with zero data migration required
- Cost centre attribution and export — integration with SAP, Oracle, and internal financial chargeback engines
- Agent Owner / Business view — run history logs, cost attribution trends, and human-in-the-loop escalation queues in Control Center
- Tier B deployment envelope — fully on-premise execution architecture; bare-metal and self-managed Kubernetes targets with zero cogward control plane dependency
Phase 2 success condition: Multiple production customers. First compliance certification complete — one framework, one vertical, validated with real auditors. Governance layer generating real policy telemetry, not speculative configuration.
Agent Lifecycle & Deployment Maturity
Pillar 5 (Estate Intelligence) reaches full — behavioural drift detection, fleet pattern intelligence, air-gapped distribution. Pillar 4 extended — snapshot-based forensic replay, data lineage. Pillar 1 extended — Tier C deployment. Sector compliance packs.
- Goal declaration and outcome tracking — declaration, progress signals, outcome attribution; from "did this agent run" to "did this agent achieve its objective"
- Task progress tracking — runtime progress visibility across long-running agent tasks
- Runtime traces linked to evaluation outcomes — execution evidence tied directly to business outcome records
- Snapshot-linked forensic replay engine — event history + retrieval context + memory state = deterministic replay input
- Data lineage through agent execution — per-output chain from response to specific data source, access control, and retrieval timestamp
- Policy sandbox over historical traces — test proposed policy changes against recorded production traces before going live
- Sector-specific compliance packs — HIPAA, MiFID II, EU AI Act Article 13, following first customer validation
- Extended memory governance — cross-session agent memory parameters with structural residency controls and retention policies
- Context provider registry — declarative, governed access to organizational context providers via the agent manifest schema
- Organisational knowledge access governance — centralized access management for Confluence, SharePoint, and proprietary data stores without data duplication
- Agent-to-agent governance — PII boundary enforcement, automated permission inheritance (deny-by-default), and multi-agent delegation audit chains
- Air-gapped intelligence distribution pipelines — one-way delivery of pre-trained behavioral models and anomaly patterns pushed via signed physical media
- Tier C deployment pipeline — air-gapped, sovereign, and highly classified defense estates; initialization of FedRAMP / IL-5 compliance engineering tracks
Phase 3 success condition: The platform is the operational system of record for the enterprise agent estate — not just the runtime for individual agents. Goals, outcomes, and lineage are visible and attributable. First defence reference customer in design partner programme.
Managed Options & Autonomous Operations
Full estate lifecycle system of record. Managed service offering (after self-hosted credibility). Governed self-improvement infrastructure. Standards contribution for manifest schema and audit event format.
- Managed operations for BYOC — operational support, upgrade guidance, health review, and deployment hardening inside the customer environment
- Dedicated managed control plane — optional cogward-operated control plane for organisations that accept managed infrastructure
- Optional cloud-managed runtime — Tier A equivalent, vendor-operated, same compliance evidence as self-hosted
- Partner-operated deployments — SI and partner-run cogward deployments for large enterprise programmes
- Long-running agent reference patterns — the durable execution engine already supports hours-long tasks; formalise as a first-class capability as the market matures toward day-long and week-long autonomous work
- Governed self-improvement research track — self-improvement only when outcomes are verifiable and the audit evidence chain is mature enough to govern them safely
- Vertical evidence packs — additional sector-specific compliance packs as the customer base expands into new regulated verticals
- Broader adapter ecosystem — community and commercial adapters for the broader framework ecosystem
Phase 4 direction: The managed service extends cogward's reach to the broader enterprise market without compromising the self-hosted story. Long-running agents and self-improvement are enabled by the platform's existing architecture — the right moment to formalise them as first-class capabilities is when the regulated enterprise base is proven and the evidence chain is mature enough to govern them safely.
Not building now
cogward governs how production agents access these systems. It does not need to own all of them.
Design partner programme
Co-design the production path.
Design partners get direct engineering access, early access to Phase 2 capabilities, and meaningful input into the roadmap sequence. The audit format, deployment runbooks, and compliance reporting templates are validated against your requirements — not hypothetical ones. This is not a waitlist. It is a working relationship.