BLOS is our proprietary AI substrate — a unified, cited knowledge base, purpose-built agents, and an orchestration layer that keeps every output grounded and every run verifiable. The ultimate equaliser: the same rigour a large team takes years to build, available from day one.
It is self-evolving: every query, review, and build cycle teaches BLOS how your context works. It adapts to its environment rather than waiting for you to prompt it differently.
What's inside
Defined roles. Own memory. No claim without a source.
Each agent operates in a specific domain — strategy, research, design, review — with its own episodic memory and a hard discipline: it either cites the source or says nothing. No generated confidence. No hallucinated conclusions.
Plans in. Working software out. Self-reviewed before a human sees it.
A second tier of agents turns approved plans into working software on a modern, reliable stack. They review their own output — for correctness, grounding, and completeness — before handing it to a human.
Routes tasks. Hands off context. Keeps every step verifiable.
Agents do not freelance. The orchestration layer decides which model and which agent handles each task, passes context between them cleanly, and maintains a verifiable record of the run against the knowledge base from start to finish.
The whole environment, packaged for your team.
The Harness is the assembled whole — knowledge base, agents, orchestration, and substrate — configured for your context. The same harness that runs BlueLeaf is what gets calibrated to your organisation when you come on board.
How it works
The brief is read, ambiguities are flagged, and a scope is locked. Nothing starts without a clear, agreed objective.
Relevant chunks are retrieved from the knowledge base. The agent builds only on what it can cite — every claim traces to a source before the draft begins.
A draft is produced and immediately reviewed by the same agent layer. Ungrounded claims are flagged and corrected before a human sees the output.
Approved outputs are handed to build agents. The full run — sources, decisions, and citations — stays verifiable against the knowledge base at every transition.
Deep Agents · Build Agents · Orchestration
The fleet runs across all four layers — each agent limited to its domain, each constrained by Cite-or-Silent. The orchestration router hands context between them and tracks the full run.
Illustrative — generic roles, no live project data.
Layers
Strategy Analyst
Deep Agents
Last task · 3 min ago
Episodic: onDeep Research
Deep Agents
Last task · now
Episodic: onBuild Agent A
Build Agents
Last task · 11 min ago
Build Agent B
Build Agents
Last task · 28 min ago
Orchestration Router
Orchestration
Last task · < 1 min ago
KB Auditor
Deep Agents
Last task · 1 hr ago
Episodic: onCitation log
Competitive landscape — Sector X
“The dominant players in Sector X have converged on a subscription model…”
Internal report · Section 4
Regulatory timeline Q3
“Draft guidelines are expected in the third quarter with a 90-day comment window…”
Policy brief · Para 2
Product differentiation signals
“Early-adopter cohorts prioritised reliability over feature breadth by a 2:1 margin…”
Research summary · Appendix A
Cite-or-Silent enforced
0 ungrounded outputs in this run
Grounded Core
Strategy, research, design, and build all read from — and write back to — the same grounded core. Every retrieved chunk carries a confidence score and a traceable source. Nothing an agent says is unmoored from a source it can name.
Illustrative — generic clusters and representative citations only.
The Run · live view
Brief locked
Grounding active
Queued
Queued
Active · Retrieve & Ground
Task assigned · retrieving grounded sources
Retrieved citations
Sector overview — internal report, section 2
Competitive signals — research brief, para 7
Market sizing — ops memo, appendix B
Run log
Brief received
Scope confirmed
KB query dispatched
5 clusters scanned
Sources ranked
Top 3 selected
Draft generation
Pending retrieval
Self-review pass
Pending
Handoff to Build Agents
Pending
Verifiable at every step
Each transition checks against the KB before continuing.
Illustrative — generic task and representative run data only.
The operating principle
Every output is grounded in your actual knowledge base — or the system says nothing at all. No hallucinated confidence. No filler to fill a word count. If BLOS cannot cite it, BLOS does not say it.
This is enforced at the orchestration layer — not as a prompt instruction, but as a structural constraint wired into every agent's output path. It is why BLOS scales without getting sloppy.
Substrate
The unified layer — knowledge base, agents, and orchestration — that your entire operation runs on.
Grounded Core
The knowledge base all agents read from and cite against. Single source of truth, enforced structurally.
Orchestration
The layer that routes tasks, hands off context, and ensures no agent freelances outside its verified scope.
BLOS · Limited early access
Tell us what you're building. We'll come back with how we'd approach it and whether BLOS is the right fit.