For companies moving AI from demo to production

The harness is the product, not the prompt.

Prompts do not ship work. A harness gives the agent its tool belt: MCPs, plugins, permissions, memory, checkpoints, evals, observability, and recovery.

Start a conversationSee the work
  • MCPs
  • permissions
  • checkpoints
  • evals
  • recovery
agent tree · deep dive
Stripe APIledger dbcheckpointL3 gateeval gateL2 gatepermissionsL1 gatebusiness system✓ shipped · recoverablereport workerstock workerrefund workerL3 · WORKERinvoice workerslack agentemail agentsynthesis agentscraper agentfulfillment agentinventory agentL2 · SUB-AGENTbilling agentL1 · LEADcomms leadL1 · LEADops leadL1 · LEADresearch leadL0 · ROOTorchestrator
01An orchestrator sits at the root
02Each lead spawns its own sub-agents
03Which spawn workers — several levels deep
04The deepest worker calls a tool
05Its output travels back up — a gate at every level
06Checked at every level — shipped, recoverable

The problem

Most AI projects fail the same way.

  1. The demo works.
  2. The first customer works.
  3. The first thousand tasks work.

Then an API times out. A model changes. A workflow runs twice. Someone refreshes.

Now you are debugging state instead of building product.

That is the problem we solve.

Proof in the field

Production
A multi-tool AI-agent SaaS, web and desktop. 12+ API integrations, hardened connectors.
Ecosystem
goalkeeper on the Claude Code marketplace, plus reaper-mcp. Servers others run.
Day one
HIPAA-adjacent. Audit logging, on-device ML, designed in.
Research
Neural codecs and diffusion models, benchmarked.
See the full record

Verifygoalkeeperreaper-mcp

What separates a system from a feature

Reliability is a design decision, not a patch.

If an agent crashes halfway through approving invoices, it resumes instead of starting over.

A comparison of two agent designs. Left, the prompt-first chain: Sub-agent runs, then Sub-agent dies, then Orchestrator retries from the top, then Duplicate charge · corrupted data — The default for agents built prompt-first.. Right, the harness chain: Checkpoint state, then Sub-agent runs, then Sub-agent dies, then Recovering agent resumes, which loops back to its checkpoint — Failure is recoverable by construction.. An execution trace records the incident: at 00.00, checkpoint written; at 00.01, tool call started; at 00.04, side effect landed; at 00.05, worker failed; at 00.06, recovery read checkpoint; at 00.06, duplicate avoided.

Prompt-first

Sub-agent runs

side effect lands — API call, row written

Sub-agent dies

mid-task, after the effect

Orchestrator retries from the top

knowing none of it

Duplicate charge · corrupted data

the agent still looks fine

The default for agents built prompt-first.

How the design holds

Checkpoint state

before any stateful operation

Sub-agent runs

idempotency token per call

skip duplicate

Sub-agent dies

failure still happens — by design

retry

Recovering agent resumes

reads checkpoint, does not repeat

resume

Failure is recoverable by construction.

Execution trace

  1. 00.00checkpoint written
  2. 00.01tool call started
  3. 00.04side effect landed
  4. 00.05worker failed
  5. 00.06recovery read checkpoint
  6. 00.06duplicate avoided
  • Idempotency tokens
  • Circuit breakers
  • Cost guards
  • budget ok

idem_key = sha256(agent_id + step + payload)

Every business is moving onto AI, the way they once moved off paper. Anyone can ship a prompt and a demo. The hard part is the architecture that keeps an agent correct, cheap, and recoverable in the real world. That is harness design — where systems are won or lost.

  1. Paper
  2. Prompt + demo
  3. Harness

MRHF-Codec vs. Meta's DAC

A neural audio codec we built and measured. High-frequency reconstruction at 6 kHz and above, against Meta's Descript Audio Codec baseline.

MetricMRHF-CodecBaseline (Meta DAC)Delta
HF-SI-SDR @ 6 kHz++11.64 dB−30.44 dB+42 dB
Bitrate11.8% lowerbaseline−11.8%
Metrics won7 / 10—7 / 10

MRHF-Codec — 11.7M-param generator, trained on 275,527 files (187.4 hrs), 240+ tests across 9 modules. Source: research project benchmark.

What we build

Three flagship capabilities.

The simple version: we make AI work in production. The full capability map has eight disciplines behind it.

Flagship01

Agent reliability & AgentOps hardening

Agents keep their place when tools fail, APIs time out, or a workflow needs a human in the loop.

Explore
Flagship02

Autonomous AI departments

A whole business function can run through coordinated agents, with approvals, limits, and audit trails.

Explore
Flagship03

Full-stack AI SaaS

Customer-facing AI products built all the way through: app, database, payments, deployment.

Explore

Five more disciplines

See all eight

How we work together

Three ways to bring us in.

01

Advisory

Discovery, impact-versus-effort scoring, an AI Scope Document. Senior judgment before you commit a budget.

02

Fractional CTO

Technical leadership for AI initiatives, no full-time hire.

03

Build

We design, build, and ship the system. Full stack, production-quality bar.

Talk to us

Your first move onto AI, or keeping a production system honest. The work we do.

Tell us where you are with AI. We will tell you plainly if we can help.

Start herehello@bonfiresystems.com