Agent orchestration: Fixing the governance gap

Fewer than one-third of decision-makers link AI to business outcomes. That statistic defines the current crisis. Agent orchestration solves the resulting governance gap. The industry has moved past simple adoption; the demand now is for measurable ROI and production safeguards. This article details how Warp evolved its Oz platform into a multi-use control plane to manage fleets like Claude Code and Codex. We examine the Tessl definition of orchestration as a distinct discipline for coordinating parallel agents. The discussion covers implementing governance gates to halt unchecked spending and preserve human oversight.

Picking a single favorite coding agent is no longer a viable strategy. Engineering leaders must balance autonomy with cost controls to prevent chaos in production. Centralized management allows organizations to transition from a "wild west" spending mindset to a data-driven operational model.

The Role of Agent Orchestration in Modern Engineering Workflows

Defining Agent Orchestration and the Multi-Use Control Plane

Agent orchestration coordinates multiple autonomous units, abandoning isolated prompt-response loops. This marks orchestration as a distinct discipline for managing parallel workflows. Unlike single-agent systems acting as individual contributors, orchestrated environments run specialized nodes for research, coding, and testing simultaneously. This architectural shift directly addresses the governance gap where fewer than one-third of decision-makers successfully link AI initiatives to business outcomes.

The multi-use control plane serves as the solution: a unified interface decoupling task execution from specific model providers. Recent platform expansions enable teams to execute various coding agents within a single terminal environment. Engineering leaders can now route queries based on cost or capability without fragmenting audit logs or context memory.

Running parallel agents introduces state management complexity that single instances avoid. The primary drawback is the operational overhead required to maintain consistent context windows across disparate models. However, separating the use from the underlying intelligence prevents vendor lock-in as model performance fluctuates. Teams gain the flexibility to switch inference engines without rebuilding their entire development infrastructure.

How Cross-Use Memory Surfaces Solutions Across Agent Runs

Cross-use memory captures successful solutions from agent tasks and surfaces them in future runs regardless of the use used. Lloyd described this memory layer as observing agents and identifying necessary information to retain. Unlike chat-based tools that discard context after interaction, this layer persists verified code patterns across different execution environments. The system currently operates in research preview with a small number of pilot customers.

Retaining state complicates distinguishing transient errors from permanent fixes. The constraint involves increased storage overhead and the need for strict validation logic to prevent propagating broken patterns. Without such gates, agent orchestration risks amplifying faults rather than compounding intelligence.

This architecture allows engineering teams to switch between different coding agents without losing institutional knowledge gained during previous sessions. Memory fidelity depends entirely on the observation layer's ability to parse semantic success. For builders, this shifts the infrastructure requirement from simple API proxying to active state management. The multi-use control plane becomes the source of truth for what works, not where code runs.

Model vs Use vs Context: Why the Use Feeds the Context

Performance relies on three distinct elements: the model, the use, and the context. Zach Lloyd asserts that performance depends on these three elements. Models provide reasoning capabilities, yet the use is the engine that feeds context into the system. Different harnesses excel at managing context windows differently, directly influencing output quality regardless of the underlying intelligence. This distinction matters as workforce interaction shifts from conversational chatbot use to autonomous agents executing multi-step workflows. As engineering teams scale, the economic model shifts from individual productivity tools to complex workflows where cost is measured by deployment reliability. The true differentiator becomes the investment in governance layers that manage data and identity.

Optimizing the use creates tension between context richness and token efficiency. A larger window allows for more historical data but increases inference latency and cost. Builders must balance these constraints based on task complexity rather than defaulting to maximum context size.

Inside the Multi-Use Control Plane Architecture

Oz as the Container Layer for Context and Audit Logs

Oz operates as a persistent container layer that separates context management and audit logs from specific model providers. Teams swap between Claude Code, Codex, or open-weight models without losing historical state or verification rules. Decoupling the execution use from the intelligence layer keeps governance policies consistent even as underlying models change. Investment in these control layers is becoming the primary differentiator for enterprise AI, shifting focus from raw model capability to the infrastructure that governs it.

Engineering teams route tasks to cost-efficient open-weight models while maintaining strict audit trails. This approach generates confidence in code generated under set rules, enabling contributors to succeed regardless of the active agent. The platform acts as a container layer for context management, memory, and audit logs, allowing teams to switch models without starting from scratch. Builders see the use not merely as a passthrough but a boundary for managing context windows. Teams should prioritize infrastructure that persists context independently of the model to avoid vendor lock-in and ensure long-term auditability. Warp open-sourced its client earlier this year and uses Oz to manage the repository, where agents handle implementation and community contributors handle direct changes. This approach serves as a test of Warp's orchestration capabilities.

Routing Claude Code, Codex, and Warp Agent via Multi-Use Control

Engineering teams route Claude Code, Codex, and Warp Agent through a unified interface to decouple task execution from specific model providers. This multi-use control plane allows operators to assign frontier models to complex reasoning tasks while directing routine refactors to cost-efficient open-weight alternatives. The architecture separates the intelligence layer from the context management required for consistent governance.

The use functions as the critical engine that feeds context into the system, determining how effectively an agent uses its context window regardless of the underlying model. Practical demonstrations show that running four parallel AI agents simultaneously within a single terminal environment notably increases throughput for distinct workstreams. This shift reflects broader industry data where nearly 98 percent of OpenAI's workforce interaction has moved from chat to autonomous workflows.

Meanwhile, the industry consensus indicates a definitive move from autocomplete and AI IDEs into engineered agent workflows, marking a quantitative shift in tool adoption. Agentic AI is evolving from being primarily a "prompting problem" to an "operational problem" requiring strong infrastructure for governance, identity, and cost management. Builders must define governance policies abstractly enough to apply to any model while remaining specific enough to catch errors. Infrastructure investment must prioritize the orchestration layer over individual model benchmarks to ensure long-term scalability.

Why Leaning on Agents Creates Pressure to Nail Orchestration

Dependencies on autonomous execution force engineering teams to prioritize orchestration mechanics over raw model intelligence. As Zach Lloyd noted, leaning on agents creates immediate pressure to nail handoff protocols and memory retention. This flexible transforms AI adoption from a prompting exercise into a rigorous operational problem requiring distributed systems thinking. Teams increasingly route specific tasks to open-weight models for cost efficiency while reserving frontier models for complex reasoning, a strategy that demands strong switching logic. Coordinating these disparate agents lacks the necessary structure for enterprise scale without a unified control plane. The shift to agentic AI involves moving from short prompt-response interactions to agents capable of running for minutes or hours autonomously. Infrastructure investment must shift from acquiring improved models to building resilient multi-agent coordination layers that enforce consistency across heterogeneous workflows.

Implementing Governance Gates and Cost Controls for AI Agents

Defining Approval Gates and Least-Privilege Permissions in Oz

Classification drives the setup of approval gates within Oz, separating low-risk website edits from high-risk data modifications. Lloyd identified a "fundamental tension" in this process, noting that companies remain hesitant to operate fully hands-off. He observed, "Making changes to our website is way lower risk than making changes to our data." This specific distinction fuels the least-privilege permissions model, pushing organizations away from a "wild west" environment toward one set by quotas, budgets, and task-specific agent usage. Engineering leaders spent the past year prioritizing AI coding tool adoption, yet the focus has pivoted sharply toward measuring ROI and stopping unchecked agents on production systems. Such a shift demands coordinated effort instead of isolated attempts to embed quality from the start.

1. Route different tasks to different agents depending on what each does best.
2. Maintain a consistent governance layer across all agents regardless of the model used.
3. Avoid locking into a single source of intelligence by supporting multiple agents simultaneously.

Stricter gates lower incident risk while simultaneously increasing latency for critical fixes. Teams balance speed against the absolute necessity of human oversight in production environments. Autonomous agents execute unintended changes across the stack without set boundaries.

Configuring Token Budgets and Audit Logs for Finance Teams

Setting token budgets per engineer begins by establishing quotas to measure ROI rather than absorbing aggregate platform costs. Finance teams increasingly demand this granularity to track spending as costs appear on balance sheets. Implementation requires mapping each developer to a specific spend limit, preventing any single user from exhausting resources on unchecked agent loops. Warp provides token efficiency per PR as a baseline for these measurements. Administrators define these constraints in the policy configuration file:

The system enables teams to use different agents for different types of tasks while keeping the governance layer consistent. These records allow finance departments to reconcile invoices against actual engineering output without manual data aggregation. Strict budgeting can stifle experimentation if limits are set too low for complex debugging sessions. Tighter controls reduce financial risk but may increase friction for engineers solving novel problems. Teams must balance cost containment with the flexibility needed for effective agent orchestration.

Risks of Unchecked Agents and Security Team Involvement

Deploying autonomous agents without oversight invites immediate production instability. Engineering leaders have shifted focus from simple adoption to preventing unchecked agents on production systems. This transition highlights a sharp tension between the drive for autonomous efficiency and the necessity of strict control. Security teams are often the most hesitant stakeholders because they bear the brunt of potential breaches. Lloyd remarked, "The security team is always involved, probably the team that's most scared." This fear stems from the architectural shift where agents execute multi-step workflows rather than simple prompt-response interactions. Unlike short conversational turns, these autonomous agents run for minutes or hours, notably expanding the attack surface.

Organizations must implement structured governance gates to mitigate these risks.

1. Invest in governance, context, identity, cost, and security layers as primary differentiators.
2. Coordinate multiple agents running in parallel through a dedicated orchestration layer.
3. Ensure the governance infrastructure is not tightly coupled to any one one agent.

Skipping these steps results in a loss of auditability and potential data corruption. Information flow becomes unverified and chaotic without a dedicated orchestrator to coordinate specialist agents. Teams must treat agent deployment as a distributed systems challenge requiring rigid operational protocols. Increased latency represents the limitation of this approach; every gate adds friction to the development lifecycle. This friction is the price of safety in an agentic environment.

Measurable ROI from Centralized Agent Management

Defining the ROI Mindset Shift from Adoption to Measurement

Quantifying specific business value now overshadows the drive for raw adoption numbers. Engineering leaders previously pushed developers toward autocomplete tools, yet rising expenses forced a pivot where CFOs demand proof of productivity. Companies have moved from a "can we get people to adopt" mindset to a "how do you measure ROI" mindset. This pressure demands replacing the "wild west" of unrestricted spending with set quotas and budgets. Teams evaluate open-weight models for specific tasks alongside frontier models to optimize inference costs. Fewer than one-third of decision-makers currently link AI initiatives to business outcomes, creating a visible gap.

Maintaining developer velocity while enforcing financial guardrails creates significant tension. Fragmented management prevents switching models as open-weight options match frontier performance at lower costs. Engineering teams face inflated inference costs that erode promised productivity gains without this flexibility. Industry analysis recommends establishing governance layers that decouple policy enforcement from the underlying intelligence source. Measurement frameworks remain stable even as optimal models change through this separation.

Tracking Token Budgets per PR as a Baseline Metric

The economic model of AI shifts from individual productivity tools costing per seat to complex workflows where "cost" measures reliability and scale of deployment. This granularity reveals when open-weight models suffice for routine refactoring, reserving expensive frontier models for complex architectural tasks.

Shifting from prompting to operations transforms AI from a productivity tool into an infrastructural cost center requiring strong infrastructure for identity and cost management. Teams configure their control plane to route tasks to different agents depending on what each does best. This baseline metric prevents the "wild west" spending patterns that trigger CFO scrutiny as teams scale.

The Governance Gap When Agents Execute Multi-Step Workflows

Data indicates that nearly all of OpenAI's workforce interaction has shifted from conversational chatbots to autonomous agents that perform extended task sequences. This transition exposes a control deficit where isolated prompts evolve into unmonitored system changes. Relying on one vendor's tool limits the ability to switch models based on cost or capability, locking teams into rigid pricing structures. Operators adopt multi-use control planes to route tasks dynamically across different models. A single model failure cascading through the entire development pipeline becomes less likely with this.

References

• Add more only when you hit a hard
• AI Agent Orchestration for Developers: The Complete 2026 Guide
• Best AI Coding Agents (June 2026): Scored Leaderboard: Codex