Agentic Development Infrastructure

The infrastructure gap under every agent deployment

Teams adopting coding agents discover that the model was the easy part. The hard part is everything a fleet of non-human contributors needs from the environment: identities that can be held accountable, a way to partition work so agents do not collide, memory that outlives a context window, budgets that stop a runaway loop, gates that encode the rules a reviewer would have applied, and evidence that a compliance team can inspect. Git, CI and a chat window do not provide these. Most teams bolt them on as prompts, scripts and conventions, which hold until the second agent or the second machine.

The layer, component by component

Nool provides that infrastructure as one system on top of the repository. Identity: every agent proposes under an agent id, and personas are durable, budget-capped identities with a charter, a model and their own memory. Coordination: leases with fencing numbers, isolated worktrees under .nool/try, and a hub server for cross-machine leases with OIDC principals. Memory: a ledger of intent, impact and scoped findings with supersession and as-of queries. Control: the envelope check, deterministic policy gates, architect and CISO holds. Budgets: daily USD limits on personas, token budgets per agent, bounded turns per run. Evidence: signed knots, attestations as refs, compliance reports on demand.

Built for regulated teams, run by your own engineers

Everything runs locally and verifies locally. Policies and roles are configuration, scaffolded by nool config init-governance and distributable as verified governance packs. Sensitive paths, auth, crypto, identity and billing, hold for a named CISO attestation before they reach the remote. nool audit report and nool audit export produce compliance evidence with framework validation. This is infrastructure you operate, on repositories you own, with agents you choose; it is the difference between governing your own build and outsourcing it, which the 8090 comparison sets out.

Measured, with the limits published

The coordination benchmark is a vendor-run benchmark, built so a skeptic can check every number. In run 9 (nool 6.14.1, synthetic agents), bare Nool landed 718.5 valid tasks per hour against 635.4 for git worktrees, spawned 77 child processes against 444 and stalled 0.2 seconds against 0.3. Governed Nool tied on throughput after correction but stalled about 0.2 seconds longer. Its memory arm gave the ledger nearly double the precision of flat retrieval. The published scope is deliberately narrow: synthetic agents, 14 of 85 commands exercised, timings that vary with load, and a lease protocol that at the time of the run cost more than git on everything but process count. Infrastructure claims should come with their limits attached, and these do.

Context economics

Infrastructure for agents also has to be economical, because token spend is the operating cost. nool context assembles what an agent needs inside an explicit budget, and with --cite returns path and line citations instead of code so the agent reads only the spans it needs; --skeletons adds AST outlines with bodies elided. nool usage bounce reports the waste pattern that matters most, context an agent compressed and then re-read in full, and persona runs can compact their working context every few turns. In the benchmark the bare coordination path spawned 77 child processes to land the workload against 444 for git worktrees; fewer processes and shorter stalls are where coordination cost shows up first, and they are measured, not asserted.

Frequently asked questions

Is this an orchestration framework like the agent SDKs?

No. SDKs build agents. Nool governs what agents may change in a repository, whichever SDK or vendor built them, and records what they did. It sits below the agent and above Git.

Do I need a server?

No. Everything works on one machine with no server. A hub server is optional, for atomic leases across machines with OIDC-based access control; Git-only teams get the same fencing as compare-and-swap refs.

How do budgets work?

Personas carry a daily USD budget. nool usage budget-set caps tokens per agent and nool usage budget-status shows where each stands. Persona runs are bounded by --max-turns, and nool usage bounce reports context an agent compressed and then re-read in full.

What does it cost?

Individuals evaluate free for 30 days or 2,000 knots with no card. Organisations license through sales for the CLI everywhere plus Mission Control. There are no self-serve paid tiers.

How does the infrastructure reduce token spend?

By orienting agents with a budgeted packet instead of a repository read, by citing spans rather than dumping code, by compacting persona context on a schedule, and by capping spend per agent and per day. nool usage analytics reports cost per line and waste ratio so the savings are measurable.