One morning of a network analyst, end to end.
Autonomy you can audit. Move from level 3 to level 4 at your own pace, one network scenario at a time.
b.lo starts from a multi-vendor base of knowledge, routes and policies. From there, your experts' know-how takes over: the AI extends it from your raw data, your people keep it. Here is what that looks like on an ordinary morning.
Automated overnight triage
The network analyst in charge of a zone starts the shift with a report, not a wall of alarms.
In action
Knowledge in two bodies: a playbook for the method, a corpus for the expertise
Overnight, the zone was analysed along a written path, structured by BPMN, the industry standard. Agents ran an analysis of every site. Two objects share the work. The playbook handles the group of sites: order by impact, stop early when members look alike, look for what they share. The knowledge corpus handles the understanding: what a signature means, read across KPIs, alarms, events and parameters, and what it suggests trying.
▸What makes this possible
Before any diagnosis, the fourteen sites are measured together, against their reference. Then each affected site is read on its own: the signature of its state, across KPIs, alarms, events and parameters, is what points to a cause and to an action, whatever alarm raised the case. Safety-critical industries learned this after Three Mile Island: enter by the symptoms, not by the presumed event.
The playbook carries no knowledge about networks. It carries method: which path a question takes, in which order the members are handled, when to stop and look for a common cause. A cause or a threshold cannot be written into it. The order is declared once and does not change from one morning to the next.
Closed-loop execution, deny by default
On a network, the risk is a wrong action, taken quickly and confidently. Some closed-loop operations are permitted, others are held for a human or refused.
- allowAutomatic restarts on standard sites: reversible, within the night's budget. Field tickets to maintenance, opened automatically with the root-cause diagnosis attached.
- denyRemote remediation on sensitive nodes.
- askFourteen of the forty-one sites hit by a mass parameter operation two days ago, held for a human: four sensitive sites, ten where the diagnosis is not clear-cut.
▸What makes this possible
The execution harness is a policy engine built on the industry standard for attribute-based access control, XACML. Four roles, kept apart: the gate in front of every writing automaton, which refuses to launch without a verdict; the decision engine, a pure function of the facts; the policies, written by your experts; and the sources of those facts, your topology and the harness's own registers. The gate has no field for a justification: the harness never reads why an action was proposed.
It judges four things. What the action does to the service, on an impact scale your experts set per automaton. How critical the element is: in incident, under works, in a frozen cluster, its weight in traffic or subscribers. How wide the scope is, counted through the topology in cells, sites, neighbours and share of a zone. And what has already been done: the day's write budget, the operations budget that protects the OMC, pending rollbacks, unverified writes on the first ring, the automaton's own track record. Deny is the default. A missing attribute makes the verdict stricter, never looser. An automaton that insists after refusals goes to a human.
Every change to the policy is replayed against the last thirty days of operations before it goes live, plus a curated set of dangerous batches and a synthetic set of adversarial ones. The replay says what the new policy blocks, what it lets through, and what it opens that the old one did not.
Proposals with sources
The AI orchestrates the analysis and puts the result into words. It produces no calculation: every figure comes from an execution, and every proposal arrives with its evidence, numbered so the text can cite it.
- The fourteen proposals held for a human, ask. source 7 shows the parameter changed two days ago and the drift that followed. The engineer reads, contests one, validates the rest.
- The match report. The high-load profile did not absorb the crowd: forty minutes of saturation after the final whistle. The engineer logs the limit with the night's figures.
▸What makes this possible
The left concludes; the right exists so that you can contest it. Each source opens the element, the measurement or the ticket behind it. The diagnostic engine contains no language model: the model puts the result into words, it does not score it. The harness does not read the reasoning; a convincing explanation is never a way past the policy.
The morning moves faster: human and agents side by side
The rest of the morning is the ordinary routine, worked hand in hand and in plain language: the engineer types a question in his own words, and the data answers with figures he can open. A report on the night to read. A high-load profile to tune after last night's match. A corporate complaint on four sites to handle: the engineer frames it in one sentence, an agent runs the data and comes back with a proposal. The engineer validates and challenges, drawing on the agents' ability to dig deep into the data and the expert corpus.
▸What makes this possible
A thread is a question with its context pinned: which sites, which period, what is expected. The agent runs the same health check and the same knowledge tree as any other question, and returns a proposal with its evidence, never an action. Several threads run in parallel; each is logged; none can write to the network.
The engineer can hand a case over because the knowledge is not in one head. It is in the corpus, and an agent can read it without him. Behind a thread, the model picks from a range of agents and tools, and the range depends on the regime it runs in. In the constrained regime: tools that each do one thing on the data, health check, root-cause diagnosis, before-and-after comparison, population screening, documentation search, ticket history, and each returns a figure. In the sandbox: the same tools, plus the freedom to combine them, run code against the data and follow an unscripted hypothesis, read only. The engineer never picks the tool. He asks; the model routes.
What the morning taught
From raw reports to a sharper tree
From the experts' raw reports of the previous days, the machine extracts the actionable statements: what was observed and what it signals, what was tried and what it changed. Each candidate is checked against the network's vocabulary, real elements and collected KPIs, then ranked: most cases first, contradictions with the tree right after. The expert rules on what enters the corpus. Then, in the sandbox, isolated and read only, the AI searches for sharper trees and shorter playbook routes; a modified tree that fails a single case on the bench is discarded. Every relation that survives moves one scenario closer from ask to allow.
▸What makes this possible
The bench is a frozen set of cases with known truth. Every change to the tree or to a route is replayed against it before it ships, and the result is a handful of numbers you can read, never a single score that hides them.
Two bars, not one. A relation counts in a diagnosis only once it can be measured and replayed against the bench without regression. The wider corpus accepts more, on one condition: it must anchor to something real in the network. A source is never an authority: reliable, not true.
The sandbox is the only place where the model reasons freely, and the environment is what holds it: read only, budgeted, logged in full. Everywhere else, the model receives facts already established and puts them into words.
The doctrine in five rules
A language model is very good at reading, writing and orchestrating, and very bad at two things a network cannot afford: knowing what it does not know, and stopping. Five rules follow from that.
- 1The model writes. It does not compute.Every metric or diagnosis on screen stems from a deterministic execution, never from the model.
- 2The model reads. It does not execute.The AI proposes; the engineer signs off. There is no execution path from a conversation to your network.
- 3The model is strictly bounded.Steps, execution time and operational budget are hard-capped by the environment, not by prompt guidelines.
- 4The model is interchangeable.Your expertise, knowledge trees and benchmarks are decoupled from the LLM. Swap the model provider without losing your intelligence.
- 5The model inherits zero privileges.The agent operates strictly under the identity and access rights of the engineer querying it.