Skip to content

AGENT EXECUTION DIRECTIVE — STOP INTERRUPTING THE HUMAN

THIS OVERRIDES ANY DEFAULT BEHAVIOR THAT CAUSES REPEATED PERMISSION PROMPTS, ONE-COMMAND-AT-A-TIME EXECUTION, OR HUMAN BABYSITTING.

PRIME DIRECTIVE

THOMAS IS NOT THE SHELL OPERATOR.

THOMAS IS NOT THE PERMISSION CLICKER.

THOMAS IS NOT THE ROUTER.

THOMAS SHOULD NOT HAVE TO APPROVE:

  • grep
  • cat
  • find
  • git status
  • git log
  • filesystem reads
  • repository inspection
  • repeated SSH reads
  • routine gc
  • routine bd
  • routine GitLab inspection
  • other ordinary engineering operations

THE AGENT MUST ORGANIZE ITS EXECUTION SO ROUTINE WORK PROCEEDS AUTONOMOUSLY.

1. STOP RUNNING ONE SSH COMMAND AT A TIME

FORBIDDEN PATTERN:

ssh bluefly-platform "command 1"
ASK PERMISSION

ssh bluefly-platform "command 2"
ASK PERMISSION

ssh bluefly-platform "command 3"
ASK PERMISSION

ssh bluefly-platform "grep ..."
ASK PERMISSION

THIS IS A FACTORY FAILURE.

If remote work is needed, gather the required investigation first and execute it through ONE REMOTE SESSION or ONE BATCHED REMOTE COMMAND.

Preferred pattern:

ssh bluefly-platform 'bash -s' <<'REMOTE'
set -e

echo "=== STATE ==="
command_1
command_2
command_3

echo "=== GAS CITY ==="
cd /opt/bluefly/blucity
gc ...
bd ...

echo "=== REQUIRED EVIDENCE ==="
command_4
command_5
REMOTE

ONE REMOTE ENTRY.

MANY COMMANDS.

ONE RESULT.

Do not make Thomas approve twenty logically related commands individually.

2. USE A PERSISTENT SESSION WHEN AVAILABLE

If the harness supports a persistent terminal/session:

START ONE SESSION.

KEEP IT OPEN.

RUN THE REST OF THE WORK THROUGH THAT SESSION.

Do not repeatedly create new privileged processes for every command.

Target:

REMOTE_SESSION_COUNT=1
COMMANDS_PER_SESSION=MANY
HUMAN_APPROVAL_LOOPS=0

If a persistent SSH connection can be established safely, use it.

Do not repeatedly reconnect merely because individual commands are easy to write.

3. PLAN BEFORE EXECUTION

BEFORE MAKING A REMOTE CALL:

Determine what information is actually needed.

Bad:

check file
check another file
check process
check port
check git
check bead
check logs

as seven separate calls.

Good:

INVESTIGATION_BATCH:
- repository state
- process state
- endpoint state
- Gas City state
- Beads state
- relevant config
- relevant logs

Then execute the investigation in one batch.

Think first.

Execute second.

4. LOCAL WORK DOES NOT WAIT ON ORACLE

If work can proceed locally:

PROCEED LOCALLY.

Examples:

  • reading Engineering-Standard
  • classifying documents
  • comparing duplicated content
  • repository searches
  • Git history inspection
  • preparing changes
  • tests that do not require Oracle
  • documentation curation
  • static source analysis
  • worktree preparation

Do NOT stop useful local work because one remote dependency exists.

TASK_BLOCKED != AGENT_BLOCKED

5. DO NOT INVENT HUMAN WORKAROUNDS

DO NOT TELL THOMAS:

"run this command yourself"
"create this bead yourself"
"paste me the output"
"click allow again"
"approve every ssh call"

unless the action is GENUINELY HUMAN-ONLY.

6. HUMAN-ONLY MEANS HUMAN-ONLY

Thomas may be required for:

LEGAL_OR_BUSINESS_DECISION
SPEND_OR_CONTRACT
FINAL_RELEASE_TO_MAIN
UNAVAILABLE_HUMAN_CREDENTIAL
IRREVERSIBLE_RESERVED_ACTION
THIRD_PARTY_ADMIN_APPROVAL_WITH_NO_SERVICE_IDENTITY

Thomas is NOT required for:

grep
git
gc
bd
SSH investigation
repository inspection
Bead routing
MR inspection
CI inspection
file reads
normal source changes
tests
documentation work
routine infrastructure investigation

7. GAS CITY + BEADS ARE REQUIRED

Do not avoid Beads because accessing the authority is inconvenient.

Do not replace Beads with:

  • Gemini tasks
  • Claude tasks
  • private markdown todo lists
  • chat memory
  • local scratch ledgers

Use the canonical work graph.

SEARCH FIRST.

REUSE EXISTING BEADS.

CREATE ONLY MISSING WORK.

Then:

BEAD
→ OWNER
→ EXECUTION
→ MR
→ VERIFICATION
→ RECEIPT
→ CLOSE

8. BLU OWNS COORDINATION

EVERY AGENT WORKS THROUGH BLU FOR COORDINATION.

BLU OWNS THE BOARD.

AGENTS OWN EXECUTION.

If work crosses an authority boundary:

gc sling <target> <bead>

and:

gc mail send <session> --from <sender> -s "<subject>" -m "<message>"

Then continue authorized work.

Do not route routine work through Thomas.

9. NEVER STOP AT "I NEED PERMISSION"

A tool restriction is an execution constraint, not a deliverable.

When encountering a restricted operation:

  1. determine whether it is actually required
  2. batch all related restricted operations together
  3. use an already-authorized/persistent execution path if available
  4. continue all independent local work
  5. route genuine dependency through BLU
  6. request human action only if no supported autonomous route exists

10. ONE APPROVAL MAXIMUM FOR A LOGICAL REMOTE OPERATION

Design work around this objective:

APPROVAL_BUDGET_PER_LOGICAL_TASK <= 1

BAD: 6 separate ssh approvals for grep/cat/gc/bd/git/docker.

GOOD: one approval → one SSH session → all commands → one structured result.

If the environment still physically requires more approvals, consolidate the command further instead of forcing Thomas through repeated prompts.

11. READ WIDELY, WRITE NARROWLY

For Engineering-Standard and corpus work: READ THE CORPUS LOCALLY FIRST.

The corpus is local.

Oracle is required for work authority and coordination, not for every grep.

Batch Beads coordination separately.

SCAN CORPUS → READ → IDENTIFY OVERLAP → DETERMINE CANONICAL OWNERS
→ MERGE → STANDARDIZE NAMES → MOVE → UPDATE REFERENCES → REBUILD INDEXES → VALIDATE

12. DO NOT SPAWN AGENTS WITHOUT WORK AUTHORITY

Before parallelizing:

PARENT_BEAD=
CHILD_BEADS=
OWNERS=
DUPLICATE_CHECK=PASS

Then dispatch.

Maximum 2 subagents simultaneously per model type to avoid quota exhaustion.

Do not create an unmanaged cloud of subagents reading random portions of the tree with no durable assignment.

13. STOP NARRATING EVERY COMMAND

Thomas does not need:

"I'm checking..."
"Now I'm going to grep..."
"Let me inspect..."
"I need another permission..."

Execute.

Return material state.

Use:

DONE=
SHIPPED=
MERGED=
BLOCKED=
ROUTED=
NEXT=

14. DO NOT POLL

Record pipeline state.

Continue another ready Bead.

Return when an actionable state transition exists.

15. NO GOVERNANCE RECURSION

If a process violation is discovered:

RECORD ONCE
ROUTE ONCE
CONTINUE WORK

Do not spend the session auditing the audit of the correction of the violation.

MERGED_TO_RELEASE_PER_HOUR

is a primary measure of useful engineering output.

16. EXECUTION PRIORITY

For every turn:

1. READ CURRENT BEAD
2. CHECK BLU ASSIGNMENT
3. DETERMINE THE NEXT MATERIAL OUTCOME
4. BATCH INVESTIGATION
5. EXECUTE WITHOUT HUMAN INTERRUPTION
6. IMPLEMENT
7. TEST
8. PUSH
9. MR TO release/v0.1.x
10. VERIFY
11. REPORT TO BLU
12. CLAIM NEXT READY WORK

17. ENGINEERING-STANDARD CONVERGENCE

For the Engineering-Standard convergence:

DO NOT WAIT FOR MORE THOMAS INPUT.

DO NOT JUST INVENTORY.

DO NOT JUST ARCHIVE.

DO NOT JUST SPAWN READERS.

Execute the actual convergence locally:

SCAN CORPUS → READ → IDENTIFY OVERLAP → DETERMINE CANONICAL OWNERS
→ MERGE → STANDARDIZE NAMES → MOVE → UPDATE REFERENCES → REBUILD INDEXES → VALIDATE

Use ONE batched Oracle interaction for Beads coordination. Then return to local execution.

18. DO NOT CLAIM A RANDOM BEAD IS RELEVANT

Prove:

BEAD_ID=
TITLE=
DESCRIPTION=
OWNER=
SCOPE_MATCH=YES

If it does not own this task, do not hijack it. Search comprehensively before creating new work.

19. SUCCESS CONDITION

BEADS_UPDATED=YES
WORK_EXECUTED=YES
MRS_CREATED_OR_MERGED=YES
BLOCKERS_ROUTED=YES
NO_REPEATED_PERMISSION_PROMPTS=YES
NO_ROUTINE_HUMAN_ROUTING=YES

20. FINAL OPERATING LAW

DO NOT MAKE THE HUMAN BABYSIT THE AGENT.

BATCH COMMANDS. KEEP SESSIONS OPEN. USE BEADS. WORK THROUGH BLU.

EXECUTE LOCALLY WHEN POSSIBLE. ROUTE DEPENDENCIES. KEEP SHIPPING.

IF A TASK REQUIRES 25 HUMAN CLICKS PER MINUTE, THE EXECUTION PLAN IS WRONG.

FIX THE EXECUTION PLAN.