Building VIOSINT started with one AI agent and finished with another. The useful lesson is not which model was better. It is how an operator knows when the current agent has stopped converging — and how to hand off the work without losing the build.
Agentic orchestration is not buying an agent. It is supervising the work.
VIOSINT is the OSINT subsystem now integrated into BDS: selected identifiers fan out across specialist tools, candidate findings are validated, and evidence is organized for review. The build plan was written before implementation: data layer, tool adapters, API, web frontend, and integration scripts.
Each phase had testable exit criteria. That matters because agentic work without phase boundaries turns into a transcript. Agentic work with phase boundaries turns into a controllable delivery process.
Hermes handled the early scaffolding well. The problem emerged when the work shifted from generating structure to wiring adapters into an existing system.
Database, credential vault, adapter shells, and early LLM-client work fit the model's strengths: precise boilerplate against a detailed spec.
The task became small but contextual: update one adapter, preserve the surrounding shape, run the test, repeat.
The agent would describe the next edit, then either not make it or touch an adjacent surface. Stronger prompt constraints did not fix the pattern.
The operational signal was specific: after repeated hard constraints not to narrate plans, the agent narrated the same intent again on the same adapter class. At that point, continuing to harden the prompt was lower-probability than switching tools.
The pivot happened after Phase 2's tests were green. That gave the next agent a clean snapshot: completed data layer and adapters behind passing tests, with API/frontend/integration still ahead.
Claude Opus picked up at Phase 3. The work pattern changed: smaller commits, conventional phase prefixes, test/doc movement alongside implementation, then Phase 4, Phase 5, standalone library extraction, and BDS-side integration.
This is the core maneuver: preserve the spec, preserve the tests, switch the worker.
Database and credential-vault scaffolding.
Adapter layer reaches a passing handoff point.
Remote agent takes over from a clean boundary.
The subsystem ships, then becomes a BDS dependency.
The same model holds very different jobs depending on how it's deployed, and each fails in its own way. Recognizing which job an agent is in tells you what to watch for.
Given a spec, the model writes the system itself — schema, adapters, integration. A wide surface and contextual edits, with a large action space. This is the job VIOSINT handed Hermes, and where non-convergence showed up as narration without action.
Given a fixed set of pre-built tools — for instance, skills exposed over an MCP server — the model only chooses which to call, and when. A narrower, bounded action space, with different failure shapes: the wrong tool, a repeated call, quiet scope drift. We didn't run VIOSINT this way — the mode is named here to mark the contrast, not to claim a run we didn't make.
In both modes the operator sets the guardrails and reads the work. The agent attempts judgment inside them; the human owns the judgment about the agent — which mode it's in, whether it's converging, and when to step in.
| Claim | What happened | Why it matters |
|---|---|---|
| different shapes | The same plan produced different output patterns under different agents. | Agents are tooling. Their work must be read and supervised. |
| handoff works | The pivot happened at a phase boundary with tests green. | Switching agents can be a controlled maneuver, not a project failure. |
| operator value | The human decision was when to stop prompting and change execution strategy. | Buying an agent does not buy orchestration judgment. |
| auditable trail | The spec, persona constraints, commit sequence, and BDS integration path remain inspectable. | The claim is evidence-based, not generic AI marketing. |
The better question is: can you tell when the AI has stopped helping, and do you act on that knowledge? The VIOSINT build is a worked example. One agent was selected for privacy and local execution. When it stopped converging, the build was handed off at a clean boundary. The system shipped.
That is the difference between agent access and agentic orchestration.
Venice Inference treats agents like execution engines inside a supervised build system: plan the phases, watch the work, recognize non-convergence, and switch tools when the evidence says to switch.
Source material: the VIOSINT build spec, the Hermes persona file with repeated hard constraints, the Gemma4 Hermes model configuration, the first viosint_core commit, the Phase 3–5 commit progression, the AGENTS.md completion banner, and the BDS-side OSINT integration notes.
— V.I. lab notes, 2026-05-27