Many coding agents. One set of rules.
Engineers choose the agent that performs best on their code, so most organizations end up running several. CompFly governs what those agents actually do, decides each call before it runs, and ships the same skills and instructions to all of them.
Rules are written once. Every tool on this list follows them, and so does the next one a team picks up.
One decision, before the command runs.
An agent packaging a quarterly export tried to copy it to a bucket outside the org. It was not attacked. It was doing its job, and the fastest path ran through customer data.
agent claude-code · repo payments-api · acting for m.okonkwo call Bash aws s3 cp ./exports/customers_2026q3.csv s3://mk-scratch-2024/ matched egress.external-bucket · pack platform-baseline v14 signals destination outside org account · payload matched customer-PII · no egress grant verdict DENY returned before the command ran returned "Use the approved export bucket in the org account. Loading skill: data-export." record entry 0x9c1f2a · signed
The denial goes back to the agent as a correction it can act on, so the work continues down the sanctioned path. And it is recorded against the agent, not the developer whose credentials it used. Downstream, the commit and the API call would both have named a person. The same holds for agents working unattended for hours, where nobody is watching a prompt and a bad call runs to completion.
From write to commit.
CompFly checks the work as it happens. The same rules cover what is unsafe and what is below standard.
When it writes code
When it runs a command
When it commits
When it calls a model
Where the built-in controls stop.
Every coding agent now ships real enterprise controls, and they work. Each one governs its own tool.
What the agent may run
Managed settings and permission rules set what commands and tools an agent can use, and cannot be overridden locally.
The rules bind that one tool. A second agent, or a script calling the model API directly, is governed by a different policy or none.
Sandboxing
Sandboxed execution and network allowlists limit what a run can touch on the machine and reach on the network.
The sandbox covers the vendor’s own runtime. Headless runs in CI, and long jobs that work unattended for hours, often sit outside it.
Sensitive data
Content exclusion and secret scanning keep named files and known secret patterns out of what the agent reads or commits.
Excluding a file is a decision made ahead of time. Whether this command, sending this payload, to this destination, should run is a decision made at the moment it happens.
Instructions and skills
Instruction files and rules give each repository its own standing guidance for the agent.
They live per repository, edited by whoever is nearest. There is no org-wide version, no way to revoke one, and no record of which version an agent was carrying when it acted.
Spend
Each tool reports its own usage and seats in its own console.
Nothing shows what a developer or a team costs across all of them, and nothing caps it. CompFly sets budgets per user and per team, applied at the model call.
Audit
Each vendor logs its own sessions in its own schema, with its own retention.
The agent runs with a developer’s credentials, so downstream logs record the person, not the agent. Four tools produce four records that do not reconcile.
What the agent may run
Managed settings and permission rules set what commands and tools an agent can use, and cannot be overridden locally.
The rules bind that one tool. A second agent, or a script calling the model API directly, is governed by a different policy or none.
Sandboxing
Sandboxed execution and network allowlists limit what a run can touch on the machine and reach on the network.
The sandbox covers the vendor’s own runtime. Headless runs in CI, and long jobs that work unattended for hours, often sit outside it.
Sensitive data
Content exclusion and secret scanning keep named files and known secret patterns out of what the agent reads or commits.
Excluding a file is a decision made ahead of time. Whether this command, sending this payload, to this destination, should run is a decision made at the moment it happens.
Instructions and skills
Instruction files and rules give each repository its own standing guidance for the agent.
They live per repository, edited by whoever is nearest. There is no org-wide version, no way to revoke one, and no record of which version an agent was carrying when it acted.
Spend
Each tool reports its own usage and seats in its own console.
Nothing shows what a developer or a team costs across all of them, and nothing caps it. CompFly sets budgets per user and per team, applied at the model call.
Audit
Each vendor logs its own sessions in its own schema, with its own retention.
The agent runs with a developer’s credentials, so downstream logs record the person, not the agent. Four tools produce four records that do not reconcile.
A team running four coding agents runs four sets of rules, four log formats, and no single record.
What the agents carry.
CompFly already sits in the path of every call. That is also how it delivers what those agents should know. One pack carries both halves: the skills and instructions they work from, and the controls for what they may do.
Versioned
Scoped
Revocable
pack platform-baseline v14 assign team:payments · repo:payments-api carries skill data-export skill incident-runbook rules house-style, review-checklist denies egress.external-bucket budget 250k tokens / dev / day rollout observe → enforce
Agents that carry the same instructions produce the same shape of work twice. A denial with a reason costs less than a merged mistake, and both cost less than a run the agent had to redo because nobody told it the house rules.
See what is running before setting a rule.
Setup takes an afternoon. Enforcement stays off until it is switched on.
Connect an existing channel
The inventory fills in
Then publish a pack
Changes apply on the next call. And the first pack is not a blank page: CompFly ships a catalog of skills and baseline rules to adopt as they are, or edit to fit how teams work.
Four log formats become one, in the SIEM the security team uses.
Every entry names the agent that acted, the rule that allowed or stopped it, and the pack version in force at the time.
Bring us one team.
Connect one team. CompFly shows every coding agent they run, what it costs, and the calls a policy would have stopped, before anything is enforced.