“I tried it on a messy support escalation, and the best part was simple: the agent stopped where our team rules said it should stop.”
Software Engineer II | Google
Agent infrastructure
Infrastructure for production AI agents.
A runtime and control plane for agents that need permissions, approvals, durable execution, observability, audit logs, and cost controls.
Agents can act through Fibonacci, but not outside policy.
Teams building agents
The people closest to the rollout want control.
“It felt less like another agent demo and more like something I could actually show security without getting laughed out of the room.”
Product Manager | Stripe
“I liked that it did the boring part well. It showed what happened, what it cost, and who still needed to approve it.”
Operations Lead | Ramp
“For HR workflows, the scary part is always access. This made the boundary feel visible instead of hidden in a prompt.”
People Systems Lead | Workday
“The audit trail changed our conversation from “can we trust this?” to “which workflow would we test first?””
Engineering Manager | Atlassian
“I tried it on a messy support escalation, and the best part was simple: the agent stopped where our team rules said it should stop.”
Software Engineer II | Google
“It felt less like another agent demo and more like something I could actually show security without getting laughed out of the room.”
Product Manager | Stripe
“I liked that it did the boring part well. It showed what happened, what it cost, and who still needed to approve it.”
Operations Lead | Ramp
“For HR workflows, the scary part is always access. This made the boundary feel visible instead of hidden in a prompt.”
People Systems Lead | Workday
“The audit trail changed our conversation from “can we trust this?” to “which workflow would we test first?””
Engineering Manager | Atlassian
Architecture
All the layers agents need to work safely.
Fibonacci sits between agent intent and company systems: governing access, approvals, execution, and observability.
Control catalog[2/4]
Control plane primitives for production agents.
Focused components for routing, governing, running, and auditing agent work.
02 / Govern
Policy + Permissions
Every requested action is checked for access, risk, secrets, and approval.
PoliciesApprovalsPermissionsSecrets
Agents can act through Fibonacci, but not outside policy.
Process
How Fibonacci works
01
Connect tools
Scoped access to Slack, Gmail, GitHub, Notion, Workspace, and internal APIs.
02
Define agent work
Use natural language, API, Python, YAML, or your existing agent framework.
03
Run through policy
Check permission, risk, approval, and execution state before action.
04
Observe and improve
Track model calls, tool calls, cost, failures, retries, and outputs.
Policy gate
Every agent action passes through policy.
Action requested
Post message to #customer-success
Agent: customer-escalation-agent
Action: post Slack summary
Policy checks
Channel allowedpass
Permission verifiedpass
No secrets detectedpass
PII needs reviewreview
External summaryreview
Decision
Human approval required
Approver: Support Lead or Admin
Use cases
Start where agentic work already hurts.
Multi-step work across tools, teams, and approval boundaries.
01
Customer escalation agents
Investigate across Gmail, Slack, Notion, and GitHub before customer-facing action.
Gmail -> Notion -> GitHub -> Slack approval
02
Weekly ops brief agents
Collect team updates and draft one executive brief with traceable sources.
Slack -> Docs -> GitHub -> Brief
03
PR and release review agents
Review risk, summarize issues, draft release notes, and notify owners.
GitHub -> Issues -> Release notes -> Slack
04
Research and monitoring agents
Watch sources, classify changes, summarize what matters, and route work.
Sources -> Classify -> Summarize -> Route
Developers
Developer-first under the hood.
Bring your own agents, tools, models, and workflows. Fibonacci provides the governed runtime.
agent_runtime.py
runtime = AgentRuntime(
tools=["slack", "gmail", "github", "notion"],
policy=Policy.require_approval(
actions=["external_message", "public_post"]
)
)PythonYAMLAPICustom toolsModel-agnostic
Cost trail
Know what each agent run costs.
Finance, ops, and engineering see the expensive parts of a workflow before it scales across the company.
Run economics
Every workflow emits cost by phase, policy decision, retry, and tool call.
[$ / run]
Total
$0.18
Every run resolves into accountable phases before it becomes budget noise.
01
Model work
61%
$0.11
02
Tool calls
22%
$0.04
03
Policy checks
11%
$0.02
04
Retries
6%
$0.01
Run total
$0.18
Policy checks
5
Budget
inside cap
Governance
Built for agents that can touch real systems.
Auditability, access control, and policy enforcement from the start.
Scoped tool permissionsHuman approval gatesAudit logsOrg policiesWorkflow registryVersion historyRollbackAPI keysSecretsCost quotas